Revert "Merge remote-tracking branch 'upstream/master' into prompt"

This reverts commit 6e42088656, reversing
changes made to 4a59bb011d.

CONTRIBUTING.md

@@ -7,7 +7,7 @@ Contributions are welcome! Here are some pointers to help you install the librar
 We recommend installing the module in editable mode with the `dev` extra requirements:
 
 ```bash
-git clone https://github.com/guillaumekln/faster-whisper.git
+git clone https://github.com/SYSTRAN/faster-whisper.git
 cd faster-whisper/
 pip install -e .[dev]
 ```

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023 Guillaume Klein
+Copyright (c) 2023 SYSTRAN
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -1,4 +1,4 @@
-[![CI](https://github.com/guillaumekln/faster-whisper/workflows/CI/badge.svg)](https://github.com/guillaumekln/faster-whisper/actions?query=workflow%3ACI) [![PyPI version](https://badge.fury.io/py/faster-whisper.svg)](https://badge.fury.io/py/faster-whisper)
+[![CI](https://github.com/SYSTRAN/faster-whisper/workflows/CI/badge.svg)](https://github.com/SYSTRAN/faster-whisper/actions?query=workflow%3ACI) [![PyPI version](https://badge.fury.io/py/faster-whisper.svg)](https://badge.fury.io/py/faster-whisper)
 
 # Faster Whisper transcription with CTranslate2
 
@@ -8,11 +8,13 @@ This implementation is up to 4 times faster than [openai/whisper](https://github
 
 ## Benchmark
 
+### Whisper
+
 For reference, here's the time and memory usage that are required to transcribe [**13 minutes**](https://www.youtube.com/watch?v=0u7tTptBo9I) of audio using different implementations:
 
 * [openai/whisper](https://github.com/openai/whisper)@[6dea21fd](https://github.com/openai/whisper/commit/6dea21fd7f7253bfe450f1e2512a0fe47ee2d258)
 * [whisper.cpp](https://github.com/ggerganov/whisper.cpp)@[3b010f9](https://github.com/ggerganov/whisper.cpp/commit/3b010f9bed9a6068609e9faf52383aea792b0362)
-* [faster-whisper](https://github.com/guillaumekln/faster-whisper)@[cce6b53e](https://github.com/guillaumekln/faster-whisper/commit/cce6b53e4554f71172dad188c45f10fb100f6e3e)
+* [faster-whisper](https://github.com/SYSTRAN/faster-whisper)@[cce6b53e](https://github.com/SYSTRAN/faster-whisper/commit/cce6b53e4554f71172dad188c45f10fb100f6e3e)
 
 ### Large-v2 model on GPU
 
@@ -36,6 +38,33 @@ For reference, here's the time and memory usage that are required to transcribe
 
 *Executed with 8 threads on a Intel(R) Xeon(R) Gold 6226R.*
 
+
+### Distil-whisper
+
+| Implementation | Precision | Beam size | Time | Gigaspeech WER |
+| --- | --- | --- | --- | --- |
+| distil-whisper/distil-large-v2 | fp16 | 4 |- | 10.36 |
+| [faster-distil-large-v2](https://huggingface.co/Systran/faster-distil-whisper-large-v2) | fp16 | 5 | - | 10.28 |
+| distil-whisper/distil-medium.en | fp16 | 4 | - | 11.21 |
+| [faster-distil-medium.en](https://huggingface.co/Systran/faster-distil-whisper-medium.en) | fp16 | 5 | - | 11.21 |
+
+*Executed with CUDA 11.4 on a NVIDIA 3090.*
+
+<details>
+<summary>testing details (click to expand)</summary>
+
+For `distil-whisper/distil-large-v2`, the WER is tested with code sample from [link](https://huggingface.co/distil-whisper/distil-large-v2#evaluation). for `faster-distil-whisper`, the WER is tested with setting:
+```python
+from faster_whisper import WhisperModel
+
+model_size = "distil-large-v2"
+# model_size = "distil-medium.en"
+# Run on GPU with FP16
+model = WhisperModel(model_size, device="cuda", compute_type="float16")
+segments, info = model.transcribe("audio.mp3", beam_size=5, language="en")
+```
+</details>
+
 ## Requirements
 
 * Python 3.8 or greater
@@ -46,28 +75,35 @@ Unlike openai-whisper, FFmpeg does **not** need to be installed on the system. T
 
 GPU execution requires the following NVIDIA libraries to be installed:
 
-* [cuBLAS for CUDA 11](https://developer.nvidia.com/cublas)
-* [cuDNN 8 for CUDA 11](https://developer.nvidia.com/cudnn)
+* [cuBLAS for CUDA 12](https://developer.nvidia.com/cublas)
+* [cuDNN 8 for CUDA 12](https://developer.nvidia.com/cudnn)
 
-There are multiple ways to install these libraries. The recommended way is described in the official NVIDIA documentation, but we also suggest other installation methods below.
+**Note**: Latest versions of `ctranslate2` support CUDA 12 only. For CUDA 11, the current workaround is downgrading to the `3.24.0` version of `ctranslate2` (This can be done with `pip install --force-reinstall ctranslate2==3.24.0` or specifying the version in a `requirements.txt`).
+
+There are multiple ways to install the NVIDIA libraries mentioned above. The recommended way is described in the official NVIDIA documentation, but we also suggest other installation methods below. 
 
 <details>
 <summary>Other installation methods (click to expand)</summary>
 
+
+**Note:** For all these methods below, keep in mind the above note regarding CUDA versions. Depending on your setup, you may need to install the _CUDA 11_ versions of libraries that correspond to the CUDA 12 libraries listed in the instructions below.
+
 #### Use Docker
 
-The libraries are installed in this official NVIDIA Docker image: `nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04`.
+The libraries (cuBLAS, cuDNN) are installed in these official NVIDIA CUDA Docker images: `nvidia/cuda:12.0.0-runtime-ubuntu20.04` or `nvidia/cuda:12.0.0-runtime-ubuntu22.04`.
 
 #### Install with `pip` (Linux only)
 
 On Linux these libraries can be installed with `pip`. Note that `LD_LIBRARY_PATH` must be set before launching Python.
 
 ```bash
-pip install nvidia-cublas-cu11 nvidia-cudnn-cu11
+pip install nvidia-cublas-cu12 nvidia-cudnn-cu12
 
 export LD_LIBRARY_PATH=`python3 -c 'import os; import nvidia.cublas.lib; import nvidia.cudnn.lib; print(os.path.dirname(nvidia.cublas.lib.__file__) + ":" + os.path.dirname(nvidia.cudnn.lib.__file__))'`
 ```
 
+**Note**: Version 9+ of `nvidia-cudnn-cu12` appears to cause issues due its reliance on cuDNN 9 (Faster-Whisper does not currently support cuDNN 9). Ensure your version of the Python package is for cuDNN 8.
+
 #### Download the libraries from Purfview's repository (Windows & Linux)
 
 Purfview's [whisper-standalone-win](https://github.com/Purfview/whisper-standalone-win) provides the required NVIDIA libraries for Windows & Linux in a [single archive](https://github.com/Purfview/whisper-standalone-win/releases/tag/libs). Decompress the archive and place the libraries in a directory included in the `PATH`.
@@ -88,19 +124,21 @@ pip install faster-whisper
 ### Install the master branch
 
 ```bash
-pip install --force-reinstall "faster-whisper @ https://github.com/guillaumekln/faster-whisper/archive/refs/heads/master.tar.gz"
+pip install --force-reinstall "faster-whisper @ https://github.com/SYSTRAN/faster-whisper/archive/refs/heads/master.tar.gz"
 ```
 
 ### Install a specific commit
 
 ```bash
-pip install --force-reinstall "faster-whisper @ https://github.com/guillaumekln/faster-whisper/archive/a4f1cc8f11433e454c3934442b5e1a4ed5e865c3.tar.gz"
+pip install --force-reinstall "faster-whisper @ https://github.com/SYSTRAN/faster-whisper/archive/a4f1cc8f11433e454c3934442b5e1a4ed5e865c3.tar.gz"
 ```
 
 </details>
 
 ## Usage
 
+### Faster-whisper
+
 ```python
 from faster_whisper import WhisperModel
 
@@ -128,6 +166,25 @@ for segment in segments:
 segments, _ = model.transcribe("audio.mp3")
 segments = list(segments)  # The transcription will actually run here.
 ```
+### Faster Distil-Whisper
+
+The Distil-Whisper checkpoints are compatible with the Faster-Whisper package. In particular, the latest [distil-large-v3](https://huggingface.co/distil-whisper/distil-large-v3)
+checkpoint is intrinsically designed to work with the Faster-Whisper transcription algorithm. The following code snippet 
+demonstrates how to run inference with distil-large-v3 on a specified audio file:
+
+```python
+from faster_whisper import WhisperModel
+
+model_size = "distil-large-v3"
+
+model = WhisperModel(model_size, device="cuda", compute_type="float16")
+segments, info = model.transcribe("audio.mp3", beam_size=5, language="en", condition_on_previous_text=False)
+
+for segment in segments:
+    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
+```
+
+For more information about the distil-large-v3 model, refer to the original [model card](https://huggingface.co/distil-whisper/distil-large-v3).
 
 ### Word-level timestamps
 
@@ -147,7 +204,7 @@ The library integrates the [Silero VAD](https://github.com/snakers4/silero-vad)
 segments, _ = model.transcribe("audio.mp3", vad_filter=True)
 ```
 
-The default behavior is conservative and only removes silence longer than 2 seconds. See the available VAD parameters and default values in the [source code](https://github.com/guillaumekln/faster-whisper/blob/master/faster_whisper/vad.py). They can be customized with the dictionary argument `vad_parameters`:
+The default behavior is conservative and only removes silence longer than 2 seconds. See the available VAD parameters and default values in the [source code](https://github.com/SYSTRAN/faster-whisper/blob/master/faster_whisper/vad.py). They can be customized with the dictionary argument `vad_parameters`:
 
 ```python
 segments, _ = model.transcribe(
@@ -170,22 +227,29 @@ logging.getLogger("faster_whisper").setLevel(logging.DEBUG)
 
 ### Going further
 
-See more model and transcription options in the [`WhisperModel`](https://github.com/guillaumekln/faster-whisper/blob/master/faster_whisper/transcribe.py) class implementation.
+See more model and transcription options in the [`WhisperModel`](https://github.com/SYSTRAN/faster-whisper/blob/master/faster_whisper/transcribe.py) class implementation.
 
 ## Community integrations
 
 Here is a non exhaustive list of open-source projects using faster-whisper. Feel free to add your project to the list!
 
+
+* [faster-whisper-server](https://github.com/fedirz/faster-whisper-server) is an OpenAI compatible server using `faster-whisper`. It's easily deployable with Docker, works with OpenAI SDKs/CLI, supports streaming, and live transcription.
+* [WhisperX](https://github.com/m-bain/whisperX) is an award-winning Python library that offers speaker diarization and accurate word-level timestamps using wav2vec2 alignment
 * [whisper-ctranslate2](https://github.com/Softcatala/whisper-ctranslate2) is a command line client based on faster-whisper and compatible with the original client from openai/whisper.
 * [whisper-diarize](https://github.com/MahmoudAshraf97/whisper-diarization) is a speaker diarization tool that is based on faster-whisper and NVIDIA NeMo.
 * [whisper-standalone-win](https://github.com/Purfview/whisper-standalone-win) Standalone CLI executables of faster-whisper for Windows, Linux & macOS. 
 * [asr-sd-pipeline](https://github.com/hedrergudene/asr-sd-pipeline) provides a scalable, modular, end to end multi-speaker speech to text solution implemented using AzureML pipelines.
 * [Open-Lyrics](https://github.com/zh-plus/Open-Lyrics) is a Python library that transcribes voice files using faster-whisper, and translates/polishes the resulting text into `.lrc` files in the desired language using OpenAI-GPT.
 * [wscribe](https://github.com/geekodour/wscribe) is a flexible transcript generation tool supporting faster-whisper, it can export word level transcript and the exported transcript then can be edited with [wscribe-editor](https://github.com/geekodour/wscribe-editor)
+* [aTrain](https://github.com/BANDAS-Center/aTrain) is a graphical user interface implementation of faster-whisper developed at the BANDAS-Center at the University of Graz for transcription and diarization in Windows ([Windows Store App](https://apps.microsoft.com/detail/atrain/9N15Q44SZNS2)) and Linux.
+* [Whisper-Streaming](https://github.com/ufal/whisper_streaming) implements real-time mode for offline Whisper-like speech-to-text models with faster-whisper as the most recommended back-end. It implements a streaming policy with self-adaptive latency based on the actual source complexity, and demonstrates the state of the art.
+* [WhisperLive](https://github.com/collabora/WhisperLive) is a nearly-live implementation of OpenAI's Whisper which uses faster-whisper as the backend to transcribe audio in real-time.
+* [Faster-Whisper-Transcriber](https://github.com/BBC-Esq/ctranslate2-faster-whisper-transcriber) is a simple but reliable voice transcriber that provides a user-friendly interface.
 
 ## Model conversion
 
-When loading a model from its size such as `WhisperModel("large-v3")`, the correspondig CTranslate2 model is automatically downloaded from the [Hugging Face Hub](https://huggingface.co/Systran).
+When loading a model from its size such as `WhisperModel("large-v3")`, the corresponding CTranslate2 model is automatically downloaded from the [Hugging Face Hub](https://huggingface.co/Systran).
 
 We also provide a script to convert any Whisper models compatible with the Transformers library. They could be the original OpenAI models or user fine-tuned models.
 

benchmark/memory_benchmark.py

@@ -0,0 +1,94 @@
+import argparse
+import time
+
+from typing import Callable
+
+import py3nvml.py3nvml as nvml
+
+from memory_profiler import memory_usage
+from utils import MyThread, get_logger, inference
+
+logger = get_logger("faster-whisper")
+parser = argparse.ArgumentParser(description="Memory benchmark")
+parser.add_argument(
+    "--gpu_memory", action="store_true", help="Measure GPU memory usage"
+)
+parser.add_argument("--device-index", type=int, default=0, help="GPU device index")
+parser.add_argument(
+    "--interval",
+    type=float,
+    default=0.5,
+    help="Interval at which measurements are collected",
+)
+args = parser.parse_args()
+device_idx = args.device_index
+interval = args.interval
+
+
+def measure_memory(func: Callable[[], None]):
+    if args.gpu_memory:
+        logger.info(
+            "Measuring maximum GPU memory usage on GPU device."
+            " Make sure to not have additional processes running on the same GPU."
+        )
+        # init nvml
+        nvml.nvmlInit()
+        handle = nvml.nvmlDeviceGetHandleByIndex(device_idx)
+        gpu_name = nvml.nvmlDeviceGetName(handle)
+        gpu_memory_limit = nvml.nvmlDeviceGetMemoryInfo(handle).total >> 20
+        gpu_power_limit = nvml.nvmlDeviceGetPowerManagementLimit(handle) / 1000.0
+        info = {"gpu_memory_usage": [], "gpu_power_usage": []}
+
+        def _get_gpu_info():
+            while True:
+                info["gpu_memory_usage"].append(
+                    nvml.nvmlDeviceGetMemoryInfo(handle).used >> 20
+                )
+                info["gpu_power_usage"].append(
+                    nvml.nvmlDeviceGetPowerUsage(handle) / 1000
+                )
+                time.sleep(interval)
+
+                if stop:
+                    break
+
+            return info
+
+        stop = False
+        thread = MyThread(_get_gpu_info, params=())
+        thread.start()
+        func()
+        stop = True
+        thread.join()
+        result = thread.get_result()
+
+        # shutdown nvml
+        nvml.nvmlShutdown()
+        max_memory_usage = max(result["gpu_memory_usage"])
+        max_power_usage = max(result["gpu_power_usage"])
+        print("GPU name: %s" % gpu_name)
+        print("GPU device index: %s" % device_idx)
+        print(
+            "Maximum GPU memory usage: %dMiB / %dMiB (%.2f%%)"
+            % (
+                max_memory_usage,
+                gpu_memory_limit,
+                (max_memory_usage / gpu_memory_limit) * 100,
+            )
+        )
+        print(
+            "Maximum GPU power usage: %dW / %dW (%.2f%%)"
+            % (
+                max_power_usage,
+                gpu_power_limit,
+                (max_power_usage / gpu_power_limit) * 100,
+            )
+        )
+    else:
+        logger.info("Measuring maximum increase of memory usage.")
+        max_usage = memory_usage(func, max_usage=True, interval=interval)
+        print("Maximum increase of RAM memory usage: %d MiB" % max_usage)
+
+
+if __name__ == "__main__":
+    measure_memory(inference)

benchmark/requirements.benchmark.txt

@@ -0,0 +1,6 @@
+transformers
+jiwer
+evaluate
+datasets
+memory_profiler
+py3nvml

benchmark/speed_benchmark.py

@@ -0,0 +1,31 @@
+import argparse
+import timeit
+
+from typing import Callable
+
+from utils import inference
+
+parser = argparse.ArgumentParser(description="Speed benchmark")
+parser.add_argument(
+    "--repeat",
+    type=int,
+    default=3,
+    help="Times an experiment will be run.",
+)
+args = parser.parse_args()
+
+
+def measure_speed(func: Callable[[], None]):
+    # as written in https://docs.python.org/3/library/timeit.html#timeit.Timer.repeat,
+    # min should be taken rather than the average
+    runtimes = timeit.repeat(
+        func,
+        repeat=args.repeat,
+        number=10,
+    )
+    print(runtimes)
+    print("Min execution time: %.3fs" % (min(runtimes) / 10.0))
+
+
+if __name__ == "__main__":
+    measure_speed(inference)

benchmark/utils.py

@@ -0,0 +1,39 @@
+import logging
+
+from threading import Thread
+from typing import Optional
+
+from faster_whisper import WhisperModel
+
+model_path = "large-v3"
+model = WhisperModel(model_path, device="cuda")
+
+
+def inference():
+    segments, info = model.transcribe("benchmark.m4a", language="fr")
+    for segment in segments:
+        print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
+
+
+def get_logger(name: Optional[str] = None) -> logging.Logger:
+    formatter = logging.Formatter("%(levelname)s: %(message)s")
+    logger = logging.getLogger(name)
+    logger.setLevel(logging.DEBUG)
+    handler = logging.StreamHandler()
+    handler.setFormatter(formatter)
+    logger.addHandler(handler)
+    return logger
+
+
+class MyThread(Thread):
+    def __init__(self, func, params):
+        super(MyThread, self).__init__()
+        self.func = func
+        self.params = params
+        self.result = None
+
+    def run(self):
+        self.result = self.func(*self.params)
+
+    def get_result(self):
+        return self.result

benchmark/wer_benchmark.py

@@ -0,0 +1,61 @@
+import argparse
+import json
+
+from datasets import load_dataset
+from evaluate import load
+from tqdm import tqdm
+from transformers.models.whisper.english_normalizer import EnglishTextNormalizer
+
+from faster_whisper import WhisperModel
+
+parser = argparse.ArgumentParser(description="WER benchmark")
+parser.add_argument(
+    "--audio_numb",
+    type=int,
+    default=None,
+    help="Specify the number of validation audio files in the dataset."
+    " Set to None to retrieve all audio files.",
+)
+args = parser.parse_args()
+
+model_path = "large-v3"
+model = WhisperModel(model_path, device="cuda")
+
+# load the dataset with streaming mode
+dataset = load_dataset("librispeech_asr", "clean", split="validation", streaming=True)
+
+# define the evaluation metric
+wer_metric = load("wer")
+normalizer = EnglishTextNormalizer(json.load(open("normalizer.json")))
+
+
+def inference(batch):
+    batch["transcription"] = []
+    for sample in batch["audio"]:
+        segments, info = model.transcribe(sample["array"], language="en")
+        batch["transcription"].append("".join([segment.text for segment in segments]))
+    batch["reference"] = batch["text"]
+    return batch
+
+
+dataset = dataset.map(function=inference, batched=True, batch_size=16)
+
+all_transcriptions = []
+all_references = []
+
+# iterate over the dataset and run inference
+for i, result in tqdm(enumerate(dataset), desc="Evaluating..."):
+    all_transcriptions.append(result["transcription"])
+    all_references.append(result["reference"])
+    if args.audio_numb and i == (args.audio_numb - 1):
+        break
+
+# normalize predictions and references
+all_transcriptions = [normalizer(transcription) for transcription in all_transcriptions]
+all_references = [normalizer(reference) for reference in all_references]
+
+# compute the WER metric
+wer = 100 * wer_metric.compute(
+    predictions=all_transcriptions, references=all_references
+)
+print("WER: %.3f" % wer)

docker/Dockerfile

@@ -0,0 +1,6 @@
+FROM nvidia/cuda:12.2.2-cudnn8-runtime-ubuntu22.04
+WORKDIR /root
+RUN apt-get update -y && apt-get install -y python3-pip
+COPY infer.py jfk.flac ./
+RUN pip3 install faster-whisper
+CMD ["python3", "infer.py"]

docker/infer.py

@@ -0,0 +1,7 @@
+from faster_whisper import WhisperModel
+
+jfk_path = "jfk.flac"
+model = WhisperModel("tiny", device="cuda")
+segments, info = model.transcribe(jfk_path, word_timestamps=True)
+for segment in segments:
+    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

faster_whisper/audio.py

@@ -102,3 +102,18 @@ def _resample_frames(frames, resampler):
     # Add None to flush the resampler.
     for frame in itertools.chain(frames, [None]):
         yield from resampler.resample(frame)
+
+
+def pad_or_trim(array, length: int, *, axis: int = -1):
+    """
+    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
+    """
+    if array.shape[axis] > length:
+        array = array.take(indices=range(length), axis=axis)
+
+    if array.shape[axis] < length:
+        pad_widths = [(0, 0)] * array.ndim
+        pad_widths[axis] = (0, length - array.shape[axis])
+        array = np.pad(array, pad_widths)
+
+    return array

faster_whisper/feature_extractor.py

@@ -142,11 +142,15 @@ class FeatureExtractor:
             data[f] = np.fft.fft(fft_signal, axis=0)[:num_fft_bins]
         return data.T
 
-    def __call__(self, waveform, padding=True):
+    def __call__(self, waveform, padding=True, chunk_length=None):
         """
         Compute the log-Mel spectrogram of the provided audio, gives similar results
         whisper's original torch implementation with 1e-5 tolerance.
         """
+        if chunk_length is not None:
+            self.n_samples = chunk_length * self.sampling_rate
+            self.nb_max_frames = self.n_samples // self.hop_length
+
         if padding:
             waveform = np.pad(waveform, [(0, self.n_samples)])
 

faster_whisper/tokenizer.py

@@ -105,6 +105,42 @@ class Tokenizer:
             [s if isinstance(s, str) else self.tokenizer.decode(s) for s in outputs]
         )
 
+    @cached_property
+    def non_speech_tokens(self) -> Tuple[int]:
+        """
+        Returns the list of tokens to suppress in order to avoid any speaker tags or non-speech
+        annotations, to prevent sampling texts that are not actually spoken in the audio, e.g.
+
+        - ♪♪♪
+        - ( SPEAKING FOREIGN LANGUAGE )
+        - [DAVID] Hey there,
+
+        keeping basic punctuations like commas, periods, question marks, exclamation points, etc.
+        """
+        symbols = list('"#()*+/:;<=>@[\\]^_`{|}~「」『』')
+        symbols += (
+            "<< >> <<< >>> -- --- -( -[ (' (\" (( )) ((( ))) [[ ]] {{ }} ♪♪ ♪♪♪".split()
+        )
+
+        # symbols that may be a single token or multiple tokens depending on the tokenizer.
+        # In case they're multiple tokens, suppress the first token, which is safe because:
+        # These are between U+2640 and U+267F miscellaneous symbols that are okay to suppress
+        # in generations, and in the 3-byte UTF-8 representation they share the first two bytes.
+        miscellaneous = set("♩♪♫♬♭♮♯")
+        assert all(0x2640 <= ord(c) <= 0x267F for c in miscellaneous)
+
+        # allow hyphens "-" and single quotes "'" between words, but not at the beginning of a word
+        result = {self.encode(" -")[0], self.encode(" '")[0]}
+        for symbol in symbols + list(miscellaneous):
+            for tokens in [
+                self.encode(symbol),
+                self.encode(" " + symbol),
+            ]:
+                if len(tokens) == 1 or symbol in miscellaneous:
+                    result.add(tokens[0])
+
+        return tuple(sorted(result))
+
     def split_to_word_tokens(
         self, tokens: List[int]
     ) -> Tuple[List[str], List[List[int]]]:

faster_whisper/transcribe.py

@@ -11,10 +11,10 @@ import ctranslate2
 import numpy as np
 import tokenizers
 
-from faster_whisper.audio import decode_audio
+from faster_whisper.audio import decode_audio, pad_or_trim
 from faster_whisper.feature_extractor import FeatureExtractor
 from faster_whisper.tokenizer import _LANGUAGE_CODES, Tokenizer
-from faster_whisper.utils import download_model, format_timestamp, get_logger
+from faster_whisper.utils import download_model, format_timestamp, get_end, get_logger
 from faster_whisper.vad import (
     SpeechTimestampsMap,
     VadOptions,
@@ -66,6 +66,10 @@ class TranscriptionOptions(NamedTuple):
     word_timestamps: bool
     prepend_punctuations: str
     append_punctuations: str
+    max_new_tokens: Optional[int]
+    clip_timestamps: Union[str, List[float]]
+    hallucination_silence_threshold: Optional[float]
+    hotwords: Optional[str]
 
 
 class TranscriptionInfo(NamedTuple):
@@ -89,12 +93,15 @@ class WhisperModel:
         num_workers: int = 1,
         download_root: Optional[str] = None,
         local_files_only: bool = False,
+        files: dict = None,
+        **model_kwargs,
     ):
         """Initializes the Whisper model.
 
         Args:
           model_size_or_path: Size of the model to use (tiny, tiny.en, base, base.en,
-            small, small.en, medium, medium.en, large-v1, large-v2, large-v3, or large), a path to a
+            small, small.en, distil-small.en, medium, medium.en, distil-medium.en, large-v1,
+            large-v2, large-v3, large, distil-large-v2 or distil-large-v3), a path to a
             converted model directory, or a CTranslate2-converted Whisper model ID from the HF Hub.
             When a size or a model ID is configured, the converted model is downloaded
             from the Hugging Face Hub.
@@ -115,10 +122,18 @@ class WhisperModel:
             are saved in the standard Hugging Face cache directory.
           local_files_only:  If True, avoid downloading the file and return the path to the
             local cached file if it exists.
+          files: Load model files from the memory. This argument is a dictionary mapping file names
+            to file contents as file-like or bytes objects. If this is set, model_path acts as an
+            identifier for this model.
         """
         self.logger = get_logger()
 
-        if os.path.isdir(model_size_or_path):
+        tokenizer_bytes, preprocessor_bytes = None, None
+        if files:
+            model_path = model_size_or_path
+            tokenizer_bytes = files.pop("tokenizer.json", None)
+            preprocessor_bytes = files.pop("preprocessor_config.json", None)
+        elif os.path.isdir(model_size_or_path):
             model_path = model_size_or_path
         else:
             model_path = download_model(
@@ -134,17 +149,20 @@ class WhisperModel:
             compute_type=compute_type,
             intra_threads=cpu_threads,
             inter_threads=num_workers,
+            files=files,
+            **model_kwargs,
         )
 
         tokenizer_file = os.path.join(model_path, "tokenizer.json")
-        if os.path.isfile(tokenizer_file):
+        if tokenizer_bytes:
+            self.hf_tokenizer = tokenizers.Tokenizer.from_buffer(tokenizer_bytes)
+        elif os.path.isfile(tokenizer_file):
             self.hf_tokenizer = tokenizers.Tokenizer.from_file(tokenizer_file)
         else:
             self.hf_tokenizer = tokenizers.Tokenizer.from_pretrained(
                 "openai/whisper-tiny" + ("" if self.model.is_multilingual else ".en")
             )
-
-        self.feat_kwargs = self._get_feature_kwargs(model_path)
+        self.feat_kwargs = self._get_feature_kwargs(model_path, preprocessor_bytes)
         self.feature_extractor = FeatureExtractor(**self.feat_kwargs)
         self.num_samples_per_token = self.feature_extractor.hop_length * 2
         self.frames_per_second = (
@@ -162,19 +180,21 @@ class WhisperModel:
         """The languages supported by the model."""
         return list(_LANGUAGE_CODES) if self.model.is_multilingual else ["en"]
 
-    def _get_feature_kwargs(self, model_path) -> dict:
-        preprocessor_config_file = os.path.join(model_path, "preprocessor_config.json")
+    def _get_feature_kwargs(self, model_path, preprocessor_bytes=None) -> dict:
         config = {}
-        if os.path.isfile(preprocessor_config_file):
-            try:
-                with open(preprocessor_config_file, "r", encoding="utf-8") as json_file:
-                    config = json.load(json_file)
-                valid_keys = signature(FeatureExtractor.__init__).parameters.keys()
-                config = {k: v for k, v in config.items() if k in valid_keys}
-            except json.JSONDecodeError as e:
-                self.logger.warning(
-                    "Could not load preprocessor_config.json: %s", str(e)
-                )
+        try:
+            config_path = os.path.join(model_path, "preprocessor_config.json")
+            if preprocessor_bytes:
+                config = json.loads(preprocessor_bytes)
+            elif os.path.isfile(config_path):
+                with open(config_path, "r", encoding="utf-8") as file:
+                    config = json.load(file)
+            else:
+                return config
+            valid_keys = signature(FeatureExtractor.__init__).parameters.keys()
+            return {k: v for k, v in config.items() if k in valid_keys}
+        except json.JSONDecodeError as e:
+            self.logger.warning("Could not load preprocessor config: %s", e)
 
         return config
 
@@ -213,6 +233,13 @@ class WhisperModel:
         append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
         vad_filter: bool = False,
         vad_parameters: Optional[Union[dict, VadOptions]] = None,
+        max_new_tokens: Optional[int] = None,
+        chunk_length: Optional[int] = None,
+        clip_timestamps: Union[str, List[float]] = "0",
+        hallucination_silence_threshold: Optional[float] = None,
+        hotwords: Optional[str] = None,
+        language_detection_threshold: Optional[float] = None,
+        language_detection_segments: int = 1,
     ) -> Tuple[Iterable[Segment], TranscriptionInfo]:
         """Transcribes an input file.
 
@@ -250,7 +277,7 @@ class WhisperModel:
           prefix: Optional text to provide as a prefix for the first window.
           suppress_blank: Suppress blank outputs at the beginning of the sampling.
           suppress_tokens: List of token IDs to suppress. -1 will suppress a default set
-            of symbols as defined in the model config.json file.
+            of symbols as defined in `tokenizer.non_speech_tokens()`
           without_timestamps: Only sample text tokens.
           max_initial_timestamp: The initial timestamp cannot be later than this.
           word_timestamps: Extract word-level timestamps using the cross-attention pattern
@@ -264,7 +291,22 @@ class WhisperModel:
             https://github.com/snakers4/silero-vad.
           vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available
             parameters and default values in the class `VadOptions`).
-
+          max_new_tokens: Maximum number of new tokens to generate per-chunk. If not set,
+            the maximum will be set by the default max_length.
+          chunk_length: The length of audio segments. If it is not None, it will overwrite the
+            default chunk_length of the FeatureExtractor.
+          clip_timestamps:
+            Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to
+             process. The last end timestamp defaults to the end of the file.
+             vad_filter will be ignored if clip_timestamps is used.
+          hallucination_silence_threshold:
+            When word_timestamps is True, skip silent periods longer than this threshold
+             (in seconds) when a possible hallucination is detected
+          hotwords:
+            Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None.
+          language_detection_threshold: If the maximum probability of the language tokens is higher
+           than this value, the language is detected.
+          language_detection_segments: Number of segments to consider for the language detection.
         Returns:
           A tuple with:
 
@@ -283,7 +325,7 @@ class WhisperModel:
             "Processing audio with duration %s", format_timestamp(duration)
         )
 
-        if vad_filter:
+        if vad_filter and clip_timestamps == "0":
             if vad_parameters is None:
                 vad_parameters = VadOptions()
             elif isinstance(vad_parameters, dict):
@@ -313,7 +355,7 @@ class WhisperModel:
         else:
             speech_chunks = None
 
-        features = self.feature_extractor(audio)
+        features = self.feature_extractor(audio, chunk_length=chunk_length)
 
         encoder_output = None
         all_language_probs = None
@@ -323,15 +365,62 @@ class WhisperModel:
                 language = "en"
                 language_probability = 1
             else:
-                segment = features[:, : self.feature_extractor.nb_max_frames]
-                encoder_output = self.encode(segment)
-                # results is a list of tuple[str, float] with language names and
-                # probabilities.
-                results = self.model.detect_language(encoder_output)[0]
-                # Parse language names to strip out markers
-                all_language_probs = [(token[2:-2], prob) for (token, prob) in results]
-                # Get top language token and probability
-                language, language_probability = all_language_probs[0]
+                if (
+                    language_detection_segments is None
+                    or language_detection_segments < 1
+                ):
+                    language_detection_segments = 1
+                start_timestamp = (
+                    float(clip_timestamps.split(",")[0])
+                    if isinstance(clip_timestamps, str)
+                    else clip_timestamps[0]
+                )
+                content_frames = (
+                    features.shape[-1] - self.feature_extractor.nb_max_frames
+                )
+                seek = (
+                    int(start_timestamp * self.frames_per_second)
+                    if start_timestamp * self.frames_per_second < content_frames
+                    else 0
+                )
+                end_frames = min(
+                    seek
+                    + self.feature_extractor.nb_max_frames
+                    * language_detection_segments,
+                    content_frames,
+                )
+                detected_language_info = {}
+                while seek <= end_frames:
+                    segment = features[
+                        :, seek : seek + self.feature_extractor.nb_max_frames
+                    ]
+                    encoder_output = self.encode(segment)
+                    # results is a list of tuple[str, float] with language names and
+                    # probabilities.
+                    results = self.model.detect_language(encoder_output)[0]
+                    # Parse language names to strip out markers
+                    all_language_probs = [
+                        (token[2:-2], prob) for (token, prob) in results
+                    ]
+                    # Get top language token and probability
+                    language, language_probability = all_language_probs[0]
+                    if (
+                        language_detection_threshold is None
+                        or language_probability > language_detection_threshold
+                    ):
+                        break
+                    detected_language_info.setdefault(language, []).append(
+                        language_probability
+                    )
+                    seek += segment.shape[-1]
+                else:
+                    # If no language detected for all segments, the majority vote of the highest
+                    # projected languages for all segments is used to determine the language.
+                    language = max(
+                        detected_language_info,
+                        key=lambda lang: len(detected_language_info[lang]),
+                    )
+                    language_probability = max(detected_language_info[language])
 
                 self.logger.info(
                     "Detected language '%s' with probability %.2f",
@@ -373,12 +462,20 @@ class WhisperModel:
             initial_prompt=initial_prompt,
             prefix=prefix,
             suppress_blank=suppress_blank,
-            suppress_tokens=get_suppressed_tokens(tokenizer, suppress_tokens),
+            suppress_tokens=(
+                get_suppressed_tokens(tokenizer, suppress_tokens)
+                if suppress_tokens
+                else suppress_tokens
+            ),
             without_timestamps=without_timestamps,
             max_initial_timestamp=max_initial_timestamp,
             word_timestamps=word_timestamps,
             prepend_punctuations=prepend_punctuations,
             append_punctuations=append_punctuations,
+            max_new_tokens=max_new_tokens,
+            clip_timestamps=clip_timestamps,
+            hallucination_silence_threshold=hallucination_silence_threshold,
+            hotwords=hotwords,
         )
 
         segments = self.generate_segments(features, tokenizer, options, encoder_output)
@@ -395,7 +492,6 @@ class WhisperModel:
             vad_options=vad_parameters,
             all_language_probs=all_language_probs,
         )
-
         return segments, info
 
     def generate_segments(
@@ -406,8 +502,35 @@ class WhisperModel:
         encoder_output: Optional[ctranslate2.StorageView] = None,
     ) -> Iterable[Segment]:
         content_frames = features.shape[-1] - self.feature_extractor.nb_max_frames
+        content_duration = float(content_frames * self.feature_extractor.time_per_frame)
+
+        if isinstance(options.clip_timestamps, str):
+            options = options._replace(
+                clip_timestamps=[
+                    float(ts)
+                    for ts in (
+                        options.clip_timestamps.split(",")
+                        if options.clip_timestamps
+                        else []
+                    )
+                ]
+            )
+        seek_points: List[int] = [
+            round(ts * self.frames_per_second) for ts in options.clip_timestamps
+        ]
+        if len(seek_points) == 0:
+            seek_points.append(0)
+        if len(seek_points) % 2 == 1:
+            seek_points.append(content_frames)
+        seek_clips: List[Tuple[int, int]] = list(
+            zip(seek_points[::2], seek_points[1::2])
+        )
+
+        punctuation = "\"'“¿([{-\"'.。,，!！?？:：”)]}、"
+
         idx = 0
-        seek = 0
+        clip_idx = 0
+        seek = seek_clips[clip_idx][0]
         all_tokens = []
         all_prompt_text = []
         prompt_reset_since = 0
@@ -421,13 +544,34 @@ class WhisperModel:
                 all_tokens.extend(options.initial_prompt)
 
         last_speech_timestamp = 0.0
-        while seek < content_frames:
+        # NOTE: This loop is obscurely flattened to make the diff readable.
+        # A later commit should turn this into a simpler nested loop.
+        # for seek_clip_start, seek_clip_end in seek_clips:
+        #     while seek < seek_clip_end
+        while clip_idx < len(seek_clips):
+            seek_clip_start, seek_clip_end = seek_clips[clip_idx]
+            if seek_clip_end > content_frames:
+                seek_clip_end = content_frames
+            if seek < seek_clip_start:
+                seek = seek_clip_start
+            if seek >= seek_clip_end:
+                clip_idx += 1
+                if clip_idx < len(seek_clips):
+                    seek = seek_clips[clip_idx][0]
+                continue
             time_offset = seek * self.feature_extractor.time_per_frame
-            segment = features[:, seek : seek + self.feature_extractor.nb_max_frames]
-            segment_size = min(
-                self.feature_extractor.nb_max_frames, content_frames - seek
+            window_end_time = float(
+                (seek + self.feature_extractor.nb_max_frames)
+                * self.feature_extractor.time_per_frame
             )
+            segment_size = min(
+                self.feature_extractor.nb_max_frames,
+                content_frames - seek,
+                seek_clip_end - seek,
+            )
+            segment = features[:, seek : seek + segment_size]
             segment_duration = segment_size * self.feature_extractor.time_per_frame
+            segment = pad_or_trim(segment, self.feature_extractor.nb_max_frames)
 
             if self.logger.isEnabledFor(logging.DEBUG):
                 self.logger.debug(
@@ -440,6 +584,7 @@ class WhisperModel:
                 previous_tokens,
                 without_timestamps=options.without_timestamps,
                 prefix=options.prefix if seek == 0 else None,
+                hotwords=options.hotwords,
             )
 
             if seek > 0 or encoder_output is None:
@@ -479,10 +624,33 @@ class WhisperModel:
             previous_seek = seek
             current_segments = []
 
+            # anomalous words are very long/short/improbable
+            def word_anomaly_score(word: dict) -> float:
+                probability = word.get("probability", 0.0)
+                duration = word["end"] - word["start"]
+                score = 0.0
+                if probability < 0.15:
+                    score += 1.0
+                if duration < 0.133:
+                    score += (0.133 - duration) * 15
+                if duration > 2.0:
+                    score += duration - 2.0
+                return score
+
+            def is_segment_anomaly(segment: Optional[dict]) -> bool:
+                if segment is None or not segment["words"]:
+                    return False
+                words = [w for w in segment["words"] if w["word"] not in punctuation]
+                words = words[:8]
+                score = sum(word_anomaly_score(w) for w in words)
+                return score >= 3 or score + 0.01 >= len(words)
+
+            def next_words_segment(segments: List[dict]) -> Optional[dict]:
+                return next((s for s in segments if s["words"]), None)
+
             single_timestamp_ending = (
                 len(tokens) >= 2
-                and tokens[-2] < tokenizer.timestamp_begin
-                and tokens[-1] >= tokenizer.timestamp_begin
+                and tokens[-2] < tokenizer.timestamp_begin <= tokens[-1]
             )
 
             consecutive_timestamps = [
@@ -565,18 +733,62 @@ class WhisperModel:
                     last_speech_timestamp=last_speech_timestamp,
                 )
 
-                word_end_timestamps = [
-                    w["end"] for s in current_segments for w in s["words"]
-                ]
-                if len(word_end_timestamps) > 0:
-                    last_speech_timestamp = word_end_timestamps[-1]
-                if not single_timestamp_ending and len(word_end_timestamps) > 0:
-                    seek_shift = round(
-                        (word_end_timestamps[-1] - time_offset) * self.frames_per_second
-                    )
+                if not single_timestamp_ending:
+                    last_word_end = get_end(current_segments)
+                    if last_word_end is not None and last_word_end > time_offset:
+                        seek = round(last_word_end * self.frames_per_second)
 
-                    if seek_shift > 0:
-                        seek = previous_seek + seek_shift
+                # skip silence before possible hallucinations
+                if options.hallucination_silence_threshold is not None:
+                    threshold = options.hallucination_silence_threshold
+
+                    # if first segment might be a hallucination, skip leading silence
+                    first_segment = next_words_segment(current_segments)
+                    if first_segment is not None and is_segment_anomaly(first_segment):
+                        gap = first_segment["start"] - time_offset
+                        if gap > threshold:
+                            seek = previous_seek + round(gap * self.frames_per_second)
+                            continue
+
+                    # skip silence before any possible hallucination that is surrounded
+                    # by silence or more hallucinations
+                    hal_last_end = last_speech_timestamp
+                    for si in range(len(current_segments)):
+                        segment = current_segments[si]
+                        if not segment["words"]:
+                            continue
+                        if is_segment_anomaly(segment):
+                            next_segment = next_words_segment(
+                                current_segments[si + 1 :]
+                            )
+                            if next_segment is not None:
+                                hal_next_start = next_segment["words"][0]["start"]
+                            else:
+                                hal_next_start = time_offset + segment_duration
+                            silence_before = (
+                                segment["start"] - hal_last_end > threshold
+                                or segment["start"] < threshold
+                                or segment["start"] - time_offset < 2.0
+                            )
+                            silence_after = (
+                                hal_next_start - segment["end"] > threshold
+                                or is_segment_anomaly(next_segment)
+                                or window_end_time - segment["end"] < 2.0
+                            )
+                            if silence_before and silence_after:
+                                seek = round(
+                                    max(time_offset + 1, segment["start"])
+                                    * self.frames_per_second
+                                )
+                                if content_duration - segment["end"] < threshold:
+                                    seek = content_frames
+                                current_segments[si:] = []
+                                break
+                        hal_last_end = segment["end"]
+
+                last_word_end = get_end(current_segments)
+                if last_word_end is not None:
+                    last_speech_timestamp = last_word_end
 
             for segment in current_segments:
                 tokens = segment["tokens"]
@@ -651,6 +863,21 @@ class WhisperModel:
         max_initial_timestamp_index = int(
             round(options.max_initial_timestamp / self.time_precision)
         )
+        if options.max_new_tokens is not None:
+            max_length = len(prompt) + options.max_new_tokens
+        else:
+            max_length = self.max_length
+
+        if max_length > self.max_length:
+            raise ValueError(
+                f"The length of the prompt is {len(prompt)}, and the `max_new_tokens` "
+                f"{max_length - len(prompt)}. Thus, the combined length of the prompt "
+                f"and `max_new_tokens` is: {max_length}. This exceeds the "
+                f"`max_length` of the Whisper model: {self.max_length}. "
+                "You should either reduce the length of your prompt, or "
+                "reduce the value of `max_new_tokens`, "
+                f"so that their combined length is less that {self.max_length}."
+            )
 
         for temperature in options.temperatures:
             if temperature > 0:
@@ -672,7 +899,7 @@ class WhisperModel:
                 length_penalty=options.length_penalty,
                 repetition_penalty=options.repetition_penalty,
                 no_repeat_ngram_size=options.no_repeat_ngram_size,
-                max_length=self.max_length,
+                max_length=max_length,
                 return_scores=True,
                 return_no_speech_prob=True,
                 suppress_blank=options.suppress_blank,
@@ -730,6 +957,8 @@ class WhisperModel:
             if (
                 options.no_speech_threshold is not None
                 and result.no_speech_prob > options.no_speech_threshold
+                and options.log_prob_threshold is not None
+                and avg_logprob < options.log_prob_threshold
             ):
                 needs_fallback = False  # silence
 
@@ -756,12 +985,19 @@ class WhisperModel:
         previous_tokens: List[int],
         without_timestamps: bool = False,
         prefix: Optional[str] = None,
+        hotwords: Optional[str] = None,
     ) -> List[int]:
         prompt = []
 
-        if previous_tokens:
+        if previous_tokens or (hotwords and not prefix):
             prompt.append(tokenizer.sot_prev)
-            prompt.extend(previous_tokens[-(self.max_length // 2 - 1) :])
+            if hotwords and not prefix:
+                hotwords_tokens = tokenizer.encode(" " + hotwords.strip())
+                if len(hotwords_tokens) >= self.max_length // 2:
+                    hotwords_tokens = hotwords_tokens[: self.max_length // 2 - 1]
+                prompt.extend(hotwords_tokens)
+            if previous_tokens:
+                prompt.extend(previous_tokens[-(self.max_length // 2 - 1) :])
 
         prompt.extend(tokenizer.sot_sequence)
 
@@ -803,6 +1039,7 @@ class WhisperModel:
         word_durations = np.array([word["end"] - word["start"] for word in alignment])
         word_durations = word_durations[word_durations.nonzero()]
         median_duration = np.median(word_durations) if len(word_durations) > 0 else 0.0
+        median_duration = min(0.7, float(median_duration))
         max_duration = median_duration * 2
 
         # hack: truncate long words at sentence boundaries.
@@ -1002,15 +1239,16 @@ def get_compression_ratio(text: str) -> float:
 
 def get_suppressed_tokens(
     tokenizer: Tokenizer,
-    suppress_tokens: Optional[List[int]],
+    suppress_tokens: Tuple[int],
 ) -> Optional[List[int]]:
-    if not suppress_tokens or -1 in suppress_tokens:
-        return suppress_tokens
+    if -1 in suppress_tokens:
+        suppress_tokens = [t for t in suppress_tokens if t >= 0]
+        suppress_tokens.extend(tokenizer.non_speech_tokens)
+    elif suppress_tokens is None or len(suppress_tokens) == 0:
+        suppress_tokens = []  # interpret empty string as an empty list
+    else:
+        assert isinstance(suppress_tokens, list), "suppress_tokens must be a list"
 
-    suppress_tokens = list(suppress_tokens)
-
-    # Ensure the following special tokens are suppressed when the user does
-    # not use the default set (-1).
     suppress_tokens.extend(
         [
             tokenizer.transcribe,
@@ -1021,7 +1259,7 @@ def get_suppressed_tokens(
         ]
     )
 
-    return sorted(set(suppress_tokens))
+    return tuple(sorted(set(suppress_tokens)))
 
 
 def merge_punctuations(alignment: List[dict], prepended: str, appended: str) -> None:

faster_whisper/utils.py

@@ -22,6 +22,10 @@ _MODELS = {
     "large-v2": "Systran/faster-whisper-large-v2",
     "large-v3": "Systran/faster-whisper-large-v3",
     "large": "Systran/faster-whisper-large-v3",
+    "distil-large-v2": "Systran/faster-distil-whisper-large-v2",
+    "distil-medium.en": "Systran/faster-distil-whisper-medium.en",
+    "distil-small.en": "Systran/faster-distil-whisper-small.en",
+    "distil-large-v3": "Systran/faster-distil-whisper-large-v3",
 }
 
 
@@ -49,9 +53,10 @@ def download_model(
     """Downloads a CTranslate2 Whisper model from the Hugging Face Hub.
 
     Args:
-      size_or_id: Size of the model to download from https://huggingface.co/guillaumekln
-        (tiny, tiny.en, base, base.en, small, small.en medium, medium.en, large-v1, large-v2,
-        large-v3, large), or a CTranslate2-converted model ID from the Hugging Face Hub
+      size_or_id: Size of the model to download from https://huggingface.co/Systran
+        (tiny, tiny.en, base, base.en, small, small.en, distil-small.en, medium, medium.en,
+        distil-medium.en, large-v1, large-v2, large-v3, large, distil-large-v2,
+        distil-large-v3), or a CTranslate2-converted model ID from the Hugging Face Hub
         (e.g. Systran/faster-whisper-large-v3).
       output_dir: Directory where the model should be saved. If not set, the model is saved in
         the cache directory.
@@ -143,3 +148,10 @@ class disabled_tqdm(tqdm):
     def __init__(self, *args, **kwargs):
         kwargs["disable"] = True
         super().__init__(*args, **kwargs)
+
+
+def get_end(segments: List[dict]) -> Optional[float]:
+    return next(
+        (w["end"] for s in reversed(segments) for w in reversed(s["words"])),
+        segments[-1]["end"] if segments else None,
+    )

faster_whisper/vad.py

@@ -1,7 +1,6 @@
 import bisect
 import functools
 import os
-import warnings
 
 from typing import List, NamedTuple, Optional
 
@@ -25,9 +24,6 @@ class VadOptions(NamedTuple):
         split aggressively just before max_speech_duration_s.
       min_silence_duration_ms: In the end of each speech chunk wait for min_silence_duration_ms
         before separating it
-      window_size_samples: Audio chunks of window_size_samples size are fed to the silero VAD model.
-        WARNING! Silero VAD models were trained using 512, 1024, 1536 samples for 16000 sample rate.
-        Values other than these may affect model performance!!
       speech_pad_ms: Final speech chunks are padded by speech_pad_ms each side
     """
 
@@ -35,7 +31,6 @@ class VadOptions(NamedTuple):
     min_speech_duration_ms: int = 250
     max_speech_duration_s: float = float("inf")
     min_silence_duration_ms: int = 2000
-    window_size_samples: int = 1024
     speech_pad_ms: int = 400
 
 
@@ -61,15 +56,8 @@ def get_speech_timestamps(
     min_speech_duration_ms = vad_options.min_speech_duration_ms
     max_speech_duration_s = vad_options.max_speech_duration_s
     min_silence_duration_ms = vad_options.min_silence_duration_ms
-    window_size_samples = vad_options.window_size_samples
+    window_size_samples = 512
     speech_pad_ms = vad_options.speech_pad_ms
-
-    if window_size_samples not in [512, 1024, 1536]:
-        warnings.warn(
-            "Unusual window_size_samples! Supported window_size_samples:\n"
-            " - [512, 1024, 1536] for 16000 sampling_rate"
-        )
-
     sampling_rate = 16000
     min_speech_samples = sampling_rate * min_speech_duration_ms / 1000
     speech_pad_samples = sampling_rate * speech_pad_ms / 1000
@@ -84,14 +72,14 @@ def get_speech_timestamps(
     audio_length_samples = len(audio)
 
     model = get_vad_model()
-    state = model.get_initial_state(batch_size=1)
+    state, context = model.get_initial_states(batch_size=1)
 
     speech_probs = []
     for current_start_sample in range(0, audio_length_samples, window_size_samples):
         chunk = audio[current_start_sample : current_start_sample + window_size_samples]
         if len(chunk) < window_size_samples:
             chunk = np.pad(chunk, (0, int(window_size_samples - len(chunk))))
-        speech_prob, state = model(chunk, state, sampling_rate)
+        speech_prob, state, context = model(chunk, state, context, sampling_rate)
         speech_probs.append(speech_prob)
 
     triggered = False
@@ -261,12 +249,12 @@ class SileroVADModel:
             sess_options=opts,
         )
 
-    def get_initial_state(self, batch_size: int):
-        h = np.zeros((2, batch_size, 64), dtype=np.float32)
-        c = np.zeros((2, batch_size, 64), dtype=np.float32)
-        return h, c
+    def get_initial_states(self, batch_size: int):
+        state = np.zeros((2, batch_size, 128), dtype=np.float32)
+        context = np.zeros((batch_size, 64), dtype=np.float32)
+        return state, context
 
-    def __call__(self, x, state, sr: int):
+    def __call__(self, x, state, context, sr: int):
         if len(x.shape) == 1:
             x = np.expand_dims(x, 0)
         if len(x.shape) > 2:
@@ -276,16 +264,15 @@ class SileroVADModel:
         if sr / x.shape[1] > 31.25:
             raise ValueError("Input audio chunk is too short")
 
-        h, c = state
+        x = np.concatenate([context, x], axis=1)
 
         ort_inputs = {
             "input": x,
-            "h": h,
-            "c": c,
+            "state": state,
             "sr": np.array(sr, dtype="int64"),
         }
 
-        out, h, c = self.session.run(None, ort_inputs)
-        state = (h, c)
+        out, state = self.session.run(None, ort_inputs)
+        context = x[..., -64:]
 
-        return out, state
+        return out, state, context

faster_whisper/version.py

@@ -1,3 +1,3 @@
 """Version information."""
 
-__version__ = "0.10.0"
+__version__ = "1.0.3"

requirements.txt

@@ -1,5 +1,5 @@
-av==10.*
-ctranslate2>=3.22,<4
+av>=11.0,<13
+ctranslate2>=4.0,<5
 huggingface_hub>=0.13
-tokenizers>=0.13,<0.16
+tokenizers>=0.13,<1
 onnxruntime>=1.14,<2

setup.py

@@ -37,7 +37,7 @@ setup(
     long_description=get_long_description(),
     long_description_content_type="text/markdown",
     author="Guillaume Klein",
-    url="https://github.com/guillaumekln/faster-whisper",
+    url="https://github.com/SYSTRAN/faster-whisper",
     classifiers=[
         "Development Status :: 4 - Beta",
         "Intended Audience :: Developers",

tests/test_transcribe.py

@@ -1,6 +1,8 @@
 import os
 
 from faster_whisper import WhisperModel, decode_audio
+from faster_whisper.tokenizer import Tokenizer
+from faster_whisper.transcribe import get_suppressed_tokens
 
 
 def test_supported_languages():
@@ -97,3 +99,109 @@ def test_stereo_diarization(data_dir):
     segments, _ = model.transcribe(right)
     transcription = "".join(segment.text for segment in segments).strip()
     assert transcription == "The horizon seems extremely distant."
+
+
+def test_suppressed_tokens_minus_1():
+    model = WhisperModel("tiny.en")
+
+    tokenizer = Tokenizer(model.hf_tokenizer, False)
+    tokens = get_suppressed_tokens(tokenizer, [-1])
+    assert tokens == (
+        1,
+        2,
+        7,
+        8,
+        9,
+        10,
+        14,
+        25,
+        26,
+        27,
+        28,
+        29,
+        31,
+        58,
+        59,
+        60,
+        61,
+        62,
+        63,
+        90,
+        91,
+        92,
+        93,
+        357,
+        366,
+        438,
+        532,
+        685,
+        705,
+        796,
+        930,
+        1058,
+        1220,
+        1267,
+        1279,
+        1303,
+        1343,
+        1377,
+        1391,
+        1635,
+        1782,
+        1875,
+        2162,
+        2361,
+        2488,
+        3467,
+        4008,
+        4211,
+        4600,
+        4808,
+        5299,
+        5855,
+        6329,
+        7203,
+        9609,
+        9959,
+        10563,
+        10786,
+        11420,
+        11709,
+        11907,
+        13163,
+        13697,
+        13700,
+        14808,
+        15306,
+        16410,
+        16791,
+        17992,
+        19203,
+        19510,
+        20724,
+        22305,
+        22935,
+        27007,
+        30109,
+        30420,
+        33409,
+        34949,
+        40283,
+        40493,
+        40549,
+        47282,
+        49146,
+        50257,
+        50357,
+        50358,
+        50359,
+        50360,
+    )
+
+
+def test_suppressed_tokens_minus_value():
+    model = WhisperModel("tiny.en")
+
+    tokenizer = Tokenizer(model.hf_tokenizer, False)
+    tokens = get_suppressed_tokens(tokenizer, [13])
+    assert tokens == (13, 50257, 50357, 50358, 50359, 50360)