Add V3 Support (#578)

* Add V3 Support

* update conversion example

---------

Co-authored-by: oscaarjs <oscar.johansson@conversy.se>

README.md

@@ -104,7 +104,7 @@ pip install --force-reinstall "faster-whisper @ https://github.com/guillaumekln/
 ```python
 from faster_whisper import WhisperModel
 
-model_size = "large-v2"
+model_size = "large-v3"
 
 # Run on GPU with FP16
 model = WhisperModel(model_size, device="cuda", compute_type="float16")
@@ -185,17 +185,17 @@ Here is a non exhaustive list of open-source projects using faster-whisper. Feel
 
 ## Model conversion
 
-When loading a model from its size such as `WhisperModel("large-v2")`, the correspondig CTranslate2 model is automatically downloaded from the [Hugging Face Hub](https://huggingface.co/guillaumekln).
+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).
 
 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.
 
-For example the command below converts the [original "large-v2" Whisper model](https://huggingface.co/openai/whisper-large-v2) and saves the weights in FP16:
+For example the command below converts the [original "large-v3" Whisper model](https://huggingface.co/openai/whisper-large-v3) and saves the weights in FP16:
 
 ```bash
 pip install transformers[torch]>=4.23
 
-ct2-transformers-converter --model openai/whisper-large-v2 --output_dir whisper-large-v2-ct2 \
+ct2-transformers-converter --model openai/whisper-large-v3 --output_dir whisper-large-v3-ct2
-    --copy_files tokenizer.json --quantization float16
+--copy_files tokenizer.json preprocessor_config.json --quantization float16
 ```
 
 * The option `--model` accepts a model name on the Hub or a path to a model directory.
@@ -207,12 +207,12 @@ Models can also be converted from the code. See the [conversion API](https://ope
 
 1. Directly load the model from a local directory:
 ```python
-model = faster_whisper.WhisperModel("whisper-large-v2-ct2")
+model = faster_whisper.WhisperModel("whisper-large-v3-ct2")
 ```
 
 2. [Upload your model to the Hugging Face Hub](https://huggingface.co/docs/transformers/model_sharing#upload-with-the-web-interface) and load it from its name:
 ```python
-model = faster_whisper.WhisperModel("username/whisper-large-v2-ct2")
+model = faster_whisper.WhisperModel("username/whisper-large-v3-ct2")
 ```
 
 ## Comparing performance against other implementations

faster_whisper/tokenizer.py

@@ -108,7 +108,7 @@ class Tokenizer:
     def split_to_word_tokens(
         self, tokens: List[int]
     ) -> Tuple[List[str], List[List[int]]]:
-        if self.language_code in {"zh", "ja", "th", "lo", "my"}:
+        if self.language_code in {"zh", "ja", "th", "lo", "my", "yue"}:
             # These languages don't typically use spaces, so it is difficult to split words
             # without morpheme analysis. Here, we instead split words at any
             # position where the tokens are decoded as valid unicode points
@@ -274,4 +274,5 @@ _LANGUAGE_CODES = (
     "yi",
     "yo",
     "zh",
+    "yue",
 )

faster_whisper/transcribe.py

@@ -1,8 +1,10 @@
 import itertools
+import json
 import logging
 import os
 import zlib
 
+from inspect import signature
 from typing import BinaryIO, Iterable, List, NamedTuple, Optional, Tuple, Union
 
 import ctranslate2
@@ -92,8 +94,8 @@ class WhisperModel:
 
         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, or large), a path to a converted
+            small, small.en, medium, medium.en, large-v1, large-v2, large-v3, or large), a path to a
-            model directory, or a CTranslate2-converted Whisper model ID from the Hugging Face Hub.
+            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.
           device: Device to use for computation ("cpu", "cuda", "auto").
@@ -142,7 +144,8 @@ class WhisperModel:
                 "openai/whisper-tiny" + ("" if self.model.is_multilingual else ".en")
             )
 
-        self.feature_extractor = FeatureExtractor()
+        self.feat_kwargs = self._get_feature_kwargs(model_path)
+        self.feature_extractor = FeatureExtractor(**self.feat_kwargs)
         self.num_samples_per_token = self.feature_extractor.hop_length * 2
         self.frames_per_second = (
             self.feature_extractor.sampling_rate // self.feature_extractor.hop_length
@@ -159,6 +162,22 @@ 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")
+        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)
+                )
+
+        return config
+
     def transcribe(
         self,
         audio: Union[str, BinaryIO, np.ndarray],

faster_whisper/utils.py

@@ -10,17 +10,18 @@ import requests
 from tqdm.auto import tqdm
 
 _MODELS = {
-    "tiny.en": "guillaumekln/faster-whisper-tiny.en",
+    "tiny.en": "Systran/faster-whisper-tiny.en",
-    "tiny": "guillaumekln/faster-whisper-tiny",
+    "tiny": "Systran/faster-whisper-tiny",
-    "base.en": "guillaumekln/faster-whisper-base.en",
+    "base.en": "Systran/faster-whisper-base.en",
-    "base": "guillaumekln/faster-whisper-base",
+    "base": "Systran/faster-whisper-base",
-    "small.en": "guillaumekln/faster-whisper-small.en",
+    "small.en": "Systran/faster-whisper-small.en",
-    "small": "guillaumekln/faster-whisper-small",
+    "small": "Systran/faster-whisper-small",
-    "medium.en": "guillaumekln/faster-whisper-medium.en",
+    "medium.en": "Systran/faster-whisper-medium.en",
-    "medium": "guillaumekln/faster-whisper-medium",
+    "medium": "Systran/faster-whisper-medium",
-    "large-v1": "guillaumekln/faster-whisper-large-v1",
+    "large-v1": "Systran/faster-whisper-large-v1",
-    "large-v2": "guillaumekln/faster-whisper-large-v2",
+    "large-v2": "Systran/faster-whisper-large-v2",
-    "large": "guillaumekln/faster-whisper-large-v2",
+    "large-v3": "Systran/faster-whisper-large-v3",
+    "large": "Systran/faster-whisper-large-v3",
 }
 
 
@@ -50,8 +51,8 @@ def download_model(
     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), or a CTranslate2-converted model ID from the Hugging Face Hub
+        large-v3, large), or a CTranslate2-converted model ID from the Hugging Face Hub
-        (e.g. guillaumekln/faster-whisper-large-v2).
+        (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.
       local_files_only:  If True, avoid downloading the file and return the path to the local
@@ -76,6 +77,7 @@ def download_model(
 
     allow_patterns = [
         "config.json",
+        "preprocessor_config.json",
         "model.bin",
         "tokenizer.json",
         "vocabulary.*",

requirements.txt

@@ -1,5 +1,5 @@
 av==10.*
-ctranslate2>=3.17,<4
+ctranslate2>=3.22,<4
 huggingface_hub>=0.13
-tokenizers>=0.13,<0.15
+tokenizers>=0.13,<0.16
 onnxruntime>=1.14,<2