Run the encoder only once for each 30-second window (#73)

faster_whisper/transcribe.py

@@ -196,14 +196,16 @@ class WhisperModel:
         duration = audio.shape[0] / self.feature_extractor.sampling_rate
         features = self.feature_extractor(audio)
 
+        whisper_encoder = WhisperEncoder(self.model)
+
         if language is None:
             if not self.model.is_multilingual:
                 language = "en"
                 language_probability = 1
             else:
                 segment = features[:, : self.feature_extractor.nb_max_frames]
-                input = get_ctranslate2_storage(segment)
+                encoder_output = whisper_encoder(0, segment)
-                results = self.model.detect_language(input)
+                results = self.model.detect_language(encoder_output)
                 language_token, language_probability = results[0][0]
                 language = language_token[2:-2]
         else:
@@ -239,7 +241,7 @@ class WhisperModel:
             append_punctuations=append_punctuations,
         )
 
-        segments = self.generate_segments(features, tokenizer, options)
+        segments = self.generate_segments(features, whisper_encoder, tokenizer, options)
 
         audio_info = AudioInfo(
             language=language,
@@ -252,6 +254,7 @@ class WhisperModel:
     def generate_segments(
         self,
         features: np.ndarray,
+        whisper_encoder: "WhisperEncoder",
         tokenizer: Tokenizer,
         options: TranscriptionOptions,
     ) -> Iterable[Segment]:
@@ -281,8 +284,10 @@ class WhisperModel:
                 prefix=options.prefix,
             )
 
+            encoder_output = whisper_encoder(seek, segment)
+
             result, avg_log_prob, temperature = self.generate_with_fallback(
-                segment, prompt, tokenizer, options
+                encoder_output, prompt, tokenizer, options
             )
 
             if options.no_speech_threshold is not None:
@@ -388,7 +393,7 @@ class WhisperModel:
                 self.add_word_timestamps(
                     current_segments,
                     tokenizer,
-                    segment,
+                    encoder_output,
                     segment_size,
                     options.prepend_punctuations,
                     options.append_punctuations,
@@ -428,12 +433,11 @@ class WhisperModel:
 
     def generate_with_fallback(
         self,
-        segment: np.ndarray,
+        encoder_output: ctranslate2.StorageView,
         prompt: List[int],
         tokenizer: Tokenizer,
         options: TranscriptionOptions,
     ) -> Tuple[ctranslate2.models.WhisperGenerationResult, float, float]:
-        features = get_ctranslate2_storage(segment)
         result = None
         avg_log_prob = None
         final_temperature = None
@@ -458,7 +462,7 @@ class WhisperModel:
 
             final_temperature = temperature
             result = self.model.generate(
-                features,
+                encoder_output,
                 [prompt],
                 length_penalty=options.length_penalty,
                 max_length=self.max_length,
@@ -529,7 +533,7 @@ class WhisperModel:
         self,
         segments: List[dict],
         tokenizer: Tokenizer,
-        mel: np.ndarray,
+        encoder_output: ctranslate2.StorageView,
         num_frames: int,
         prepend_punctuations: str,
         append_punctuations: str,
@@ -543,7 +547,9 @@ class WhisperModel:
         ]
 
         text_tokens = list(itertools.chain.from_iterable(text_tokens_per_segment))
-        alignment = self.find_alignment(tokenizer, text_tokens, mel, num_frames)
+        alignment = self.find_alignment(
+            tokenizer, text_tokens, encoder_output, num_frames
+        )
         merge_punctuations(alignment, prepend_punctuations, append_punctuations)
 
         time_offset = (
@@ -585,7 +591,7 @@ class WhisperModel:
         self,
         tokenizer: Tokenizer,
         text_tokens: List[int],
-        mel: np.ndarray,
+        encoder_output: ctranslate2.StorageView,
         num_frames: int,
         median_filter_width: int = 7,
     ) -> List[dict]:
@@ -593,7 +599,7 @@ class WhisperModel:
             return []
 
         result = self.model.align(
-            get_ctranslate2_storage(mel),
+            encoder_output,
             tokenizer.sot_sequence,
             [text_tokens],
             num_frames,
@@ -646,9 +652,39 @@ class WhisperModel:
         ]
 
 
+class WhisperEncoder:
+    """Helper class to cache and reuse the encoder output."""
+
+    def __init__(self, model: ctranslate2.models.Whisper):
+        self.model = model
+
+        # When the model is running on multiple GPUs, the encoder output should be moved
+        # to the CPU since we don't know which GPU will handle the next job.
+        self.cache_on_cpu = len(model.device_index) > 1
+
+        self.last_seek = -1
+        self.last_output = None
+
+    def __call__(self, seek: int, features: np.ndarray) -> ctranslate2.StorageView:
+        if self.last_seek == seek:
+            return self.last_output
+
+        features = np.expand_dims(features, 0)
+        features = get_ctranslate2_storage(features)
+
+        output = self.model.encode(features, to_cpu=self.cache_on_cpu)
+
+        if self.last_output is not None:
+            del self.last_output
+
+        self.last_seek = seek
+        self.last_output = output
+
+        return output
+
+
 def get_ctranslate2_storage(segment: np.ndarray) -> ctranslate2.StorageView:
     segment = np.ascontiguousarray(segment)
-    segment = np.expand_dims(segment, 0)
     segment = ctranslate2.StorageView.from_array(segment)
     return segment
 

requirements.txt

@@ -1,3 +1,3 @@
 av==10.*
-ctranslate2>=3.9,<4
+ctranslate2>=3.10,<4
 tokenizers==0.13.*