Merge pull request #64 from makaveli10/update_faster_whisper

Update faster whisper.

requirements/server.txt

@@ -1,5 +1,5 @@
 PyAudio
-faster-whisper==0.6.0
+faster-whisper==0.9.0
 --extra-index-url https://download.pytorch.org/whl/cu111
 torch==1.10.1
 torchaudio==0.10.1

whisper_live/__version__.py

@@ -1 +1 @@
-__version__="0.0.7"
+__version__="0.0.8"

whisper_live/server.py

@@ -13,7 +13,6 @@ import torch
 import numpy as np
 import time
 from whisper_live.transcriber import WhisperModel
-from whisper_live.vad import VoiceActivityDetection
 
 
 class TranscriptionServer:
@@ -35,8 +34,6 @@ class TranscriptionServer:
 
     def __init__(self):
         # voice activity detection model
-        self.vad_model = VoiceActivityDetection()
-        self.vad_threshold = 0.4
 
         self.clients = {}
         self.websockets = {}
@@ -115,15 +112,6 @@ class TranscriptionServer:
                 frame_data = websocket.recv()
                 frame_np = np.frombuffer(frame_data, dtype=np.float32)
 
-                try:
-                    speech_prob = self.vad_model(torch.from_numpy(frame_np.copy()), self.RATE).item()
-                    if speech_prob < self.vad_threshold:
-                        continue
-
-                except Exception as e:
-                    logging.error(e)
-                    return
-
                 self.clients[websocket].add_frames(frame_np)
 
                 elapsed_time = time.time() - self.clients_start_time[websocket]
@@ -322,25 +310,6 @@ class ServeClient:
             Exception: If there is an issue with audio processing or WebSocket communication.
 
         """
-        # detect language
-        if self.language is None:
-            # wait for 30s of audio
-            while self.frames_np is None or self.frames_np.shape[0] < 30*self.RATE:
-                time.sleep(1)
-            input_bytes = self.frames_np[-30*self.RATE:].copy()
-            self.frames_np = None
-            duration = input_bytes.shape[0] / self.RATE
-
-            self.language, lang_prob = self.transcriber.transcribe(
-                    input_bytes, 
-                    initial_prompt=None,
-                    language=self.language,
-                    task=self.task
-                )
-            logging.info(f"Detected language {self.language} with probability {lang_prob}")
-            self.websocket.send(json.dumps(
-                {"uid": self.client_uid, "language": self.language, "language_prob": lang_prob}))
-
         while True:
             if self.exit:
                 logging.info("Exiting speech to text thread")
@@ -362,20 +331,27 @@ class ServeClient:
                 continue
             try:
                 input_sample = input_bytes.copy()
-                # set previous complete segment as initial prompt
-                if len(self.text) and self.text[-1] != '': 
-                    initial_prompt = self.text[-1]
-                else: 
-                    initial_prompt = None
                 
                 # whisper transcribe with prompt
-                result = self.transcriber.transcribe(
+                result, info = self.transcriber.transcribe(
                     input_sample, 
-                    initial_prompt=initial_prompt,
+                    initial_prompt=None,
                     language=self.language,
-                    task=self.task
+                    task=self.task,
+                    vad_filter=True,
+                    vad_parameters={"threshold": 0.5}
                 )
 
+                if self.language is None:
+                    if info.language_probability > 0.5:
+                        self.language = info.language
+                        logging.info(f"Detected language {self.language} with probability {info.language_probability}")
+                        self.websocket.send(json.dumps(
+                            {"uid": self.client_uid, "language": self.language, "language_prob": info.language_probability}))
+                    else:
+                        # detect language again
+                        continue
+
                 if len(result):
                     self.t_start = None
                     last_segment = self.update_segments(result, duration)
@@ -385,16 +361,6 @@ class ServeClient:
                         segments = self.transcript[-self.send_last_n_segments:]
                     if last_segment is not None:
                         segments = segments + [last_segment]                    
-                    
-                    try:
-                        self.websocket.send(
-                            json.dumps({
-                                "uid": self.client_uid,
-                                "segments": segments
-                            })
-                        )
-                    except Exception as e:
-                        logging.error(f"[ERROR]: {e}")
                 else:
                     # show previous output if there is pause i.e. no output from whisper
                     segments = []
@@ -419,6 +385,7 @@ class ServeClient:
                     )
                 except Exception as e:
                     logging.error(f"[ERROR]: {e}")
+
             except Exception as e:
                 logging.error(f"[ERROR]: {e}")
                 time.sleep(0.01)

whisper_live/transcriber.py

@@ -4,7 +4,6 @@ import itertools
 import logging
 import os
 import zlib
-import logging
 
 from typing import BinaryIO, Iterable, List, NamedTuple, Optional, Tuple, Union
 
@@ -14,21 +13,16 @@ import tokenizers
 
 from faster_whisper.audio import decode_audio
 from faster_whisper.feature_extractor import FeatureExtractor
-from faster_whisper.tokenizer import Tokenizer
+from faster_whisper.tokenizer import _LANGUAGE_CODES, Tokenizer
-from faster_whisper.utils import download_model, format_timestamp
+from faster_whisper.utils import download_model, format_timestamp, get_logger
 from faster_whisper.vad import (
     SpeechTimestampsMap,
+    VadOptions,
     collect_chunks,
     get_speech_timestamps,
 )
 
 
-# implement logger not available in faster_whisper==0.4.1
-def get_logger():
-    """Returns the module logger."""
-    return logging.getLogger("faster_whisper")
-
-
 class Word(NamedTuple):
     start: float
     end: float
@@ -37,18 +31,17 @@ class Word(NamedTuple):
 
 
 class Segment(NamedTuple):
+    id: int
+    seek: int
     start: float
     end: float
     text: str
-    words: Optional[List[Word]]
+    tokens: List[int]
-    avg_log_prob: float
+    temperature: float
+    avg_logprob: float
+    compression_ratio: float
     no_speech_prob: float
-
+    words: Optional[List[Word]]
-
-class AudioInfo(NamedTuple):
-    language: str
-    language_probability: float
-    duration: float
 
 
 class TranscriptionOptions(NamedTuple):
@@ -56,12 +49,15 @@ class TranscriptionOptions(NamedTuple):
     best_of: int
     patience: float
     length_penalty: float
+    repetition_penalty: float
+    no_repeat_ngram_size: int
     log_prob_threshold: Optional[float]
     no_speech_threshold: Optional[float]
     compression_ratio_threshold: Optional[float]
     condition_on_previous_text: bool
+    prompt_reset_on_temperature: float
     temperatures: List[float]
-    initial_prompt: Optional[str]
+    initial_prompt: Optional[Union[str, Iterable[int]]]
     prefix: Optional[str]
     suppress_blank: bool
     suppress_tokens: Optional[List[int]]
@@ -72,6 +68,16 @@ class TranscriptionOptions(NamedTuple):
     append_punctuations: str
 
 
+class TranscriptionInfo(NamedTuple):
+    language: str
+    language_probability: float
+    duration: float
+    duration_after_vad: float
+    all_language_probs: Optional[List[Tuple[str, float]]]
+    transcription_options: TranscriptionOptions
+    vad_options: VadOptions
+
+
 class WhisperModel:
     def __init__(
         self,
@@ -82,14 +88,15 @@ class WhisperModel:
         cpu_threads: int = 0,
         num_workers: int = 1,
         download_root: Optional[str] = None,
-        local_files_only: bool = True,
+        local_files_only: bool = False,
     ):
         """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, or large-v2) or a path to a converted
+            small, small.en, medium, medium.en, large-v1, large-v2, or large), a path to a converted
-            model directory. When a size is configured, the converted model is downloaded
+            model directory, or a CTranslate2-converted Whisper model ID from the Hugging Face 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").
           device_index: Device ID to use.
@@ -104,8 +111,10 @@ class WhisperModel:
             having multiple workers enables true parallelism when running the model
             (concurrent calls to self.model.generate() will run in parallel).
             This can improve the global throughput at the cost of increased memory usage.
-          download_root: Directory where the model should be saved. If not set, the model
+          download_root: Directory where the models should be saved. If not set, the models
-            is saved in the standard Hugging Face cache directory.
+            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.
         """
         self.logger = get_logger()
 
@@ -147,6 +156,11 @@ class WhisperModel:
         self.time_precision = 0.02
         self.max_length = 448
 
+    @property
+    def supported_languages(self) -> List[str]:
+        """The languages supported by the model."""
+        return list(_LANGUAGE_CODES) if self.model.is_multilingual else ["en"]
+
     def transcribe(
         self,
         audio: Union[str, BinaryIO, np.ndarray],
@@ -156,6 +170,8 @@ class WhisperModel:
         best_of: int = 5,
         patience: float = 1,
         length_penalty: float = 1,
+        repetition_penalty: float = 1,
+        no_repeat_ngram_size: int = 0,
         temperature: Union[float, List[float], Tuple[float, ...]] = [
             0.0,
             0.2,
@@ -168,7 +184,8 @@ class WhisperModel:
         log_prob_threshold: Optional[float] = -1.0,
         no_speech_threshold: Optional[float] = 0.6,
         condition_on_previous_text: bool = True,
-        initial_prompt: Optional[str] = None,
+        prompt_reset_on_temperature: float = 0.5,
+        initial_prompt: Optional[Union[str, Iterable[int]]] = None,
         prefix: Optional[str] = None,
         suppress_blank: bool = True,
         suppress_tokens: Optional[List[int]] = [-1],
@@ -178,8 +195,8 @@ class WhisperModel:
         prepend_punctuations: str = "\"'“¿([{-",
         append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
         vad_filter: bool = False,
-        vad_parameters: Optional[dict] = None,
+        vad_parameters: Optional[Union[dict, VadOptions]] = None,
-    ) -> Tuple[Iterable[Segment], AudioInfo]:
+    ) -> Tuple[Iterable[Segment], TranscriptionInfo]:
         """Transcribes an input file.
 
         Arguments:
@@ -192,6 +209,9 @@ class WhisperModel:
           best_of: Number of candidates when sampling with non-zero temperature.
           patience: Beam search patience factor.
           length_penalty: Exponential length penalty constant.
+          repetition_penalty: Penalty applied to the score of previously generated tokens
+            (set > 1 to penalize).
+          no_repeat_ngram_size: Prevent repetitions of ngrams with this size (set 0 to disable).
           temperature: Temperature for sampling. It can be a tuple of temperatures,
             which will be successively used upon failures according to either
             `compression_ratio_threshold` or `log_prob_threshold`.
@@ -206,7 +226,10 @@ class WhisperModel:
             as a prompt for the next window; disabling may make the text inconsistent across
             windows, but the model becomes less prone to getting stuck in a failure loop,
             such as repetition looping or timestamps going out of sync.
-          initial_prompt: Optional text to provide as a prompt for the first window.
+          prompt_reset_on_temperature: Resets prompt if temperature is above this value.
+            Arg has effect only if condition_on_previous_text is True.
+          initial_prompt: Optional text string or iterable of token ids to provide as a
+            prompt for the first window.
           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
@@ -222,14 +245,14 @@ class WhisperModel:
           vad_filter: Enable the voice activity detection (VAD) to filter out parts of the audio
             without speech. This step is using the Silero VAD model
             https://github.com/snakers4/silero-vad.
-          vad_parameters: Dictionary of Silero VAD parameters (see available parameters and
+          vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available
-            default values in the function `get_speech_timestamps`).
+            parameters and default values in the class `VadOptions`).
 
         Returns:
           A tuple with:
 
             - a generator over transcribed segments
-            - an instance of AudioInfo
+            - an instance of TranscriptionInfo
         """
         sampling_rate = self.feature_extractor.sampling_rate
 
@@ -237,19 +260,24 @@ class WhisperModel:
             audio = decode_audio(audio, sampling_rate=sampling_rate)
 
         duration = audio.shape[0] / sampling_rate
+        duration_after_vad = duration
 
         self.logger.info(
             "Processing audio with duration %s", format_timestamp(duration)
         )
 
         if vad_filter:
-            vad_parameters = {} if vad_parameters is None else vad_parameters
+            if vad_parameters is None:
-            speech_chunks = get_speech_timestamps(audio, **vad_parameters)
+                vad_parameters = VadOptions()
+            elif isinstance(vad_parameters, dict):
+                vad_parameters = VadOptions(**vad_parameters)
+            speech_chunks = get_speech_timestamps(audio, vad_parameters)
             audio = collect_chunks(audio, speech_chunks)
+            duration_after_vad = audio.shape[0] / sampling_rate
 
             self.logger.info(
                 "VAD filter removed %s of audio",
-                format_timestamp(duration - (audio.shape[0] / sampling_rate)),
+                format_timestamp(duration - duration_after_vad),
             )
 
             if self.logger.isEnabledFor(logging.DEBUG):
@@ -271,6 +299,7 @@ class WhisperModel:
         features = self.feature_extractor(audio)
 
         encoder_output = None
+        all_language_probs = None
 
         if language is None:
             if not self.model.is_multilingual:
@@ -279,17 +308,27 @@ class WhisperModel:
             else:
                 segment = features[:, : self.feature_extractor.nb_max_frames]
                 encoder_output = self.encode(segment)
-                results = self.model.detect_language(encoder_output)
+                # results is a list of tuple[str, float] with language names and
-                language_token, language_probability = results[0][0]
+                # probabilities.
-                language = language_token[2:-2]
+                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]
 
                 self.logger.info(
                     "Detected language '%s' with probability %.2f",
                     language,
                     language_probability,
                 )
-                return language, language_probability
         else:
+            if not self.model.is_multilingual and language != "en":
+                self.logger.warning(
+                    "The current model is English-only but the language parameter is set to '%s'; "
+                    "using 'en' instead." % language
+                )
+                language = "en"
+
             language_probability = 1
 
         tokenizer = Tokenizer(
@@ -304,10 +343,13 @@ class WhisperModel:
             best_of=best_of,
             patience=patience,
             length_penalty=length_penalty,
+            repetition_penalty=repetition_penalty,
+            no_repeat_ngram_size=no_repeat_ngram_size,
             log_prob_threshold=log_prob_threshold,
             no_speech_threshold=no_speech_threshold,
             compression_ratio_threshold=compression_ratio_threshold,
             condition_on_previous_text=condition_on_previous_text,
+            prompt_reset_on_temperature=prompt_reset_on_temperature,
             temperatures=(
                 temperature if isinstance(temperature, (list, tuple)) else [temperature]
             ),
@@ -327,13 +369,17 @@ class WhisperModel:
         if speech_chunks:
             segments = restore_speech_timestamps(segments, speech_chunks, sampling_rate)
 
-        audio_info = AudioInfo(
+        info = TranscriptionInfo(
             language=language,
             language_probability=language_probability,
             duration=duration,
+            duration_after_vad=duration_after_vad,
+            transcription_options=options,
+            vad_options=vad_parameters,
+            all_language_probs=all_language_probs,
         )
 
-        return segments
+        return segments, info
 
     def generate_segments(
         self,
@@ -343,14 +389,20 @@ class WhisperModel:
         encoder_output: Optional[ctranslate2.StorageView] = None,
     ) -> Iterable[Segment]:
         content_frames = features.shape[-1] - self.feature_extractor.nb_max_frames
+        idx = 0
         seek = 0
         all_tokens = []
         prompt_reset_since = 0
 
         if options.initial_prompt is not None:
+            if isinstance(options.initial_prompt, str):
                 initial_prompt = " " + options.initial_prompt.strip()
                 initial_prompt_tokens = tokenizer.encode(initial_prompt)
                 all_tokens.extend(initial_prompt_tokens)
+            else:
+                all_tokens.extend(options.initial_prompt)
+
+        last_speech_timestamp = 0.0
         all_segments = []
         while seek < content_frames:
             time_offset = seek * self.feature_extractor.time_per_frame
@@ -373,12 +425,15 @@ class WhisperModel:
                 prefix=options.prefix if seek == 0 else None,
             )
 
-            if encoder_output is None:
+            if seek > 0 or encoder_output is None:
                 encoder_output = self.encode(segment)
 
-            result, avg_log_prob, temperature = self.generate_with_fallback(
+            (
-                encoder_output, prompt, tokenizer, options
+                result,
-            )
+                avg_logprob,
+                temperature,
+                compression_ratio,
+            ) = self.generate_with_fallback(encoder_output, prompt, tokenizer, options)
 
             if options.no_speech_threshold is not None:
                 # no voice activity check
@@ -386,7 +441,7 @@ class WhisperModel:
 
                 if (
                     options.log_prob_threshold is not None
-                    and avg_log_prob > options.log_prob_threshold
+                    and avg_logprob > options.log_prob_threshold
                 ):
                     # don't skip if the logprob is high enough, despite the no_speech_prob
                     should_skip = False
@@ -482,9 +537,6 @@ class WhisperModel:
 
                 seek += segment_size
 
-            if not options.condition_on_previous_text or temperature > 0.5:
-                prompt_reset_since = len(all_tokens)
-
             if options.word_timestamps:
                 self.add_word_timestamps(
                     current_segments,
@@ -493,12 +545,14 @@ class WhisperModel:
                     segment_size,
                     options.prepend_punctuations,
                     options.append_punctuations,
+                    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
@@ -507,8 +561,6 @@ class WhisperModel:
                     if seek_shift > 0:
                         seek = previous_seek + seek_shift
 
-            encoder_output = None
-            
             for segment in current_segments:
                 tokens = segment["tokens"]
                 text = tokenizer.decode(tokens)
@@ -517,19 +569,38 @@ class WhisperModel:
                     continue
 
                 all_tokens.extend(tokens)
+                idx += 1
 
                 all_segments.append(Segment(
+                    id=idx,
+                    seek=seek,
                     start=segment["start"],
                     end=segment["end"],
                     text=text,
+                    tokens=tokens,
+                    temperature=temperature,
+                    avg_logprob=avg_logprob,
+                    compression_ratio=compression_ratio,
+                    no_speech_prob=result.no_speech_prob,
                     words=(
                         [Word(**word) for word in segment["words"]]
                         if options.word_timestamps
                         else None
                         ),
-                    avg_log_prob=avg_log_prob,
-                    no_speech_prob=result.no_speech_prob,
                 ))
+
+            if (
+                not options.condition_on_previous_text
+                or temperature > options.prompt_reset_on_temperature
+            ):
+                if options.condition_on_previous_text:
+                    self.logger.debug(
+                        "Reset prompt. prompt_reset_on_temperature threshold is met %f > %f",
+                        temperature,
+                        options.prompt_reset_on_temperature,
+                    )
+
+                prompt_reset_since = len(all_tokens)
         return all_segments
 
     def encode(self, features: np.ndarray) -> ctranslate2.StorageView:
@@ -548,10 +619,10 @@ class WhisperModel:
         prompt: List[int],
         tokenizer: Tokenizer,
         options: TranscriptionOptions,
-    ) -> Tuple[ctranslate2.models.WhisperGenerationResult, float, float]:
+    ) -> Tuple[ctranslate2.models.WhisperGenerationResult, float, float, float]:
-        result = None
+        decode_result = None
-        avg_log_prob = None
+        all_results = []
-        final_temperature = None
+        below_cr_threshold_results = []
 
         max_initial_timestamp_index = int(
             round(options.max_initial_timestamp / self.time_precision)
@@ -571,11 +642,12 @@ class WhisperModel:
                     "patience": options.patience,
                 }
 
-            final_temperature = temperature
             result = self.model.generate(
                 encoder_output,
                 [prompt],
                 length_penalty=options.length_penalty,
+                repetition_penalty=options.repetition_penalty,
+                no_repeat_ngram_size=options.no_repeat_ngram_size,
                 max_length=self.max_length,
                 return_scores=True,
                 return_no_speech_prob=True,
@@ -589,18 +661,24 @@ class WhisperModel:
 
             # Recover the average log prob from the returned score.
             seq_len = len(tokens)
-            cum_log_prob = result.scores[0] * (seq_len**options.length_penalty)
+            cum_logprob = result.scores[0] * (seq_len**options.length_penalty)
-            avg_log_prob = cum_log_prob / (seq_len + 1)
+            avg_logprob = cum_logprob / (seq_len + 1)
 
             text = tokenizer.decode(tokens).strip()
             compression_ratio = get_compression_ratio(text)
 
+            decode_result = (
+                result,
+                avg_logprob,
+                temperature,
+                compression_ratio,
+            )
+            all_results.append(decode_result)
+
             needs_fallback = False
 
-            if (
+            if options.compression_ratio_threshold is not None:
-                options.compression_ratio_threshold is not None
+                if compression_ratio > options.compression_ratio_threshold:
-                and compression_ratio > options.compression_ratio_threshold
-            ):
                     needs_fallback = True  # too repetitive
 
                     self.logger.debug(
@@ -609,24 +687,37 @@ class WhisperModel:
                         compression_ratio,
                         options.compression_ratio_threshold,
                     )
+                else:
+                    below_cr_threshold_results.append(decode_result)
 
             if (
                 options.log_prob_threshold is not None
-                and avg_log_prob < options.log_prob_threshold
+                and avg_logprob < options.log_prob_threshold
             ):
                 needs_fallback = True  # average log probability is too low
 
                 self.logger.debug(
                     "Log probability threshold is not met with temperature %.1f (%f < %f)",
                     temperature,
-                    avg_log_prob,
+                    avg_logprob,
                     options.log_prob_threshold,
                 )
 
+            if (
+                options.no_speech_threshold is not None
+                and result.no_speech_prob > options.no_speech_threshold
+            ):
+                needs_fallback = False  # silence
+
             if not needs_fallback:
                 break
+        else:
+            # all failed, select the result with the highest average log probability
+            decode_result = max(
+                below_cr_threshold_results or all_results, key=lambda x: x[1]
+            )
 
-        return result, avg_log_prob, final_temperature
+        return decode_result
 
     def get_prompt(
         self,
@@ -650,6 +741,8 @@ class WhisperModel:
             prefix_tokens = tokenizer.encode(" " + prefix.strip())
             if len(prefix_tokens) >= self.max_length // 2:
                 prefix_tokens = prefix_tokens[: self.max_length // 2 - 1]
+            if not without_timestamps:
+                prompt.append(tokenizer.timestamp_begin)
             prompt.extend(prefix_tokens)
 
         return prompt
@@ -662,7 +755,8 @@ class WhisperModel:
         num_frames: int,
         prepend_punctuations: str,
         append_punctuations: str,
-    ):
+        last_speech_timestamp: float,
+    ) -> None:
         if len(segments) == 0:
             return
 
@@ -675,6 +769,24 @@ class WhisperModel:
         alignment = self.find_alignment(
             tokenizer, text_tokens, encoder_output, num_frames
         )
+        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
+        max_duration = median_duration * 2
+
+        # hack: truncate long words at sentence boundaries.
+        # a better segmentation algorithm based on VAD should be able to replace this.
+        if len(word_durations) > 0:
+            sentence_end_marks = ".。!！?？"
+            # ensure words at sentence boundaries
+            # are not longer than twice the median word duration.
+            for i in range(1, len(alignment)):
+                if alignment[i]["end"] - alignment[i]["start"] > max_duration:
+                    if alignment[i]["word"] in sentence_end_marks:
+                        alignment[i]["end"] = alignment[i]["start"] + max_duration
+                    elif alignment[i - 1]["word"] in sentence_end_marks:
+                        alignment[i]["start"] = alignment[i]["end"] - max_duration
+
         merge_punctuations(alignment, prepend_punctuations, append_punctuations)
 
         time_offset = (
@@ -705,11 +817,52 @@ class WhisperModel:
                 saved_tokens += len(timing["tokens"])
                 word_index += 1
 
+            # hack: truncate long words at segment boundaries.
+            # a better segmentation algorithm based on VAD should be able to replace this.
             if len(words) > 0:
-                # adjust the segment-level timestamps based on the word-level timestamps
+                # ensure the first and second word after a pause is not longer than
+                # twice the median word duration.
+                if words[0]["end"] - last_speech_timestamp > median_duration * 4 and (
+                    words[0]["end"] - words[0]["start"] > max_duration
+                    or (
+                        len(words) > 1
+                        and words[1]["end"] - words[0]["start"] > max_duration * 2
+                    )
+                ):
+                    if (
+                        len(words) > 1
+                        and words[1]["end"] - words[1]["start"] > max_duration
+                    ):
+                        boundary = max(
+                            words[1]["end"] / 2, words[1]["end"] - max_duration
+                        )
+                        words[0]["end"] = words[1]["start"] = boundary
+                    words[0]["start"] = max(0, words[0]["end"] - max_duration)
+
+                # prefer the segment-level start timestamp if the first word is too long.
+                if (
+                    segment["start"] < words[0]["end"]
+                    and segment["start"] - 0.5 > words[0]["start"]
+                ):
+                    words[0]["start"] = max(
+                        0, min(words[0]["end"] - median_duration, segment["start"])
+                    )
+                else:
                     segment["start"] = words[0]["start"]
+
+                # prefer the segment-level end timestamp if the last word is too long.
+                if (
+                    segment["end"] > words[-1]["start"]
+                    and segment["end"] + 0.5 < words[-1]["end"]
+                ):
+                    words[-1]["end"] = max(
+                        words[-1]["start"] + median_duration, segment["end"]
+                    )
+                else:
                     segment["end"] = words[-1]["end"]
 
+                last_speech_timestamp = segment["end"]
+
             segment["words"] = words
 
     def find_alignment(
@@ -741,6 +894,8 @@ class WhisperModel:
             text_tokens + [tokenizer.eot]
         )
         word_boundaries = np.pad(np.cumsum([len(t) for t in word_tokens[:-1]]), (1, 0))
+        if len(word_boundaries) <= 1:
+            return []
 
         jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(bool)
         jump_times = time_indices[jumps] / self.tokens_per_second
@@ -751,22 +906,6 @@ class WhisperModel:
             for i, j in zip(word_boundaries[:-1], word_boundaries[1:])
         ]
 
-        # hack: ensure the first and second word is not longer than twice the median word duration.
-        # a better segmentation algorithm based on VAD should be able to replace this.
-        word_durations = end_times - start_times
-        word_durations = word_durations[word_durations.nonzero()]
-        if len(word_durations) > 0:
-            median_duration = np.median(word_durations)
-            max_duration = median_duration * 2
-            if len(word_durations) >= 2 and word_durations[1] > max_duration:
-                boundary = max(end_times[2] / 2, end_times[2] - max_duration)
-                end_times[0] = start_times[1] = boundary
-            if (
-                len(word_durations) >= 1
-                and end_times[0] - start_times[0] > max_duration
-            ):
-                start_times[0] = max(0, end_times[0] - max_duration)
-
         return [
             dict(
                 word=word, tokens=tokens, start=start, end=end, probability=probability
@@ -792,7 +931,8 @@ def restore_speech_timestamps(
             words = []
             for word in segment.words:
                 # Ensure the word start and end times are resolved to the same chunk.
-                chunk_index = ts_map.get_chunk_index(word.start)
+                middle = (word.start + word.end) / 2
+                chunk_index = ts_map.get_chunk_index(middle)
                 word = word._replace(
                     start=ts_map.get_original_time(word.start, chunk_index),
                     end=ts_map.get_original_time(word.end, chunk_index),
@@ -811,7 +951,7 @@ def restore_speech_timestamps(
                 end=ts_map.get_original_time(segment.end),
             )
 
-        yield segment
+    return segments
 
 
 def get_ctranslate2_storage(segment: np.ndarray) -> ctranslate2.StorageView:
@@ -825,7 +965,10 @@ def get_compression_ratio(text: str) -> float:
     return len(text_bytes) / len(zlib.compress(text_bytes))
 
 
-def get_suppressed_tokens(tokenizer, suppress_tokens):
+def get_suppressed_tokens(
+    tokenizer: Tokenizer,
+    suppress_tokens: Optional[List[int]],
+) -> Optional[List[int]]:
     if not suppress_tokens or -1 in suppress_tokens:
         return suppress_tokens
 
@@ -846,7 +989,7 @@ def get_suppressed_tokens(tokenizer, suppress_tokens):
     return sorted(set(suppress_tokens))
 
 
-def merge_punctuations(alignment: List[dict], prepended: str, appended: str):
+def merge_punctuations(alignment: List[dict], prepended: str, appended: str) -> None:
     # merge prepended punctuations
     i = len(alignment) - 2
     j = len(alignment) - 1

whisper_live/vad.py

@@ -1,115 +0,0 @@
-# original: https://github.com/snakers4/silero-vad/blob/master/utils_vad.py
-
-import os
-import subprocess
-import torch
-import numpy as np
-import onnxruntime
-
-
-class VoiceActivityDetection():
-
-    def __init__(self, force_onnx_cpu=True):
-        path = self.download()
-        opts = onnxruntime.SessionOptions()
-        opts.log_severity_level = 3
-
-        opts.inter_op_num_threads = 1
-        opts.intra_op_num_threads = 1
-
-        if force_onnx_cpu and 'CPUExecutionProvider' in onnxruntime.get_available_providers():
-            self.session = onnxruntime.InferenceSession(path, providers=['CPUExecutionProvider'], sess_options=opts)
-        else:
-            self.session = onnxruntime.InferenceSession(path, providers=['CUDAExecutionProvider'], sess_options=opts)
-
-
-        self.reset_states()
-        self.sample_rates = [8000, 16000]
-
-    def _validate_input(self, x, sr: int):
-        if x.dim() == 1:
-            x = x.unsqueeze(0)
-        if x.dim() > 2:
-            raise ValueError(f"Too many dimensions for input audio chunk {x.dim()}")
-
-        if sr != 16000 and (sr % 16000 == 0):
-            step = sr // 16000
-            x = x[:,::step]
-            sr = 16000
-
-        if sr not in self.sample_rates:
-            raise ValueError(f"Supported sampling rates: {self.sample_rates} (or multiply of 16000)")
-
-        if sr / x.shape[1] > 31.25:
-            raise ValueError("Input audio chunk is too short")
-
-        return x, sr
-
-    def reset_states(self, batch_size=1):
-        self._h = np.zeros((2, batch_size, 64)).astype('float32')
-        self._c = np.zeros((2, batch_size, 64)).astype('float32')
-        self._last_sr = 0
-        self._last_batch_size = 0
-
-    def __call__(self, x, sr: int):
-
-        x, sr = self._validate_input(x, sr)
-        batch_size = x.shape[0]
-
-        if not self._last_batch_size:
-            self.reset_states(batch_size)
-        if (self._last_sr) and (self._last_sr != sr):
-            self.reset_states(batch_size)
-        if (self._last_batch_size) and (self._last_batch_size != batch_size):
-            self.reset_states(batch_size)
-
-        if sr in [8000, 16000]:
-            ort_inputs = {'input': x.numpy(), 'h': self._h, 'c': self._c, 'sr': np.array(sr, dtype='int64')}
-            ort_outs = self.session.run(None, ort_inputs)
-            out, self._h, self._c = ort_outs
-        else:
-            raise ValueError()
-
-        self._last_sr = sr
-        self._last_batch_size = batch_size
-
-        out = torch.tensor(out)
-        return out
-
-    def audio_forward(self, x, sr: int, num_samples: int = 512):
-        outs = []
-        x, sr = self._validate_input(x, sr)
-
-        if x.shape[1] % num_samples:
-            pad_num = num_samples - (x.shape[1] % num_samples)
-            x = torch.nn.functional.pad(x, (0, pad_num), 'constant', value=0.0)
-
-        self.reset_states(x.shape[0])
-        for i in range(0, x.shape[1], num_samples):
-            wavs_batch = x[:, i:i+num_samples]
-            out_chunk = self.__call__(wavs_batch, sr)
-            outs.append(out_chunk)
-
-        stacked = torch.cat(outs, dim=1)
-        return stacked.cpu()
-
-    @staticmethod
-    def download(model_url="https://github.com/snakers4/silero-vad/raw/master/files/silero_vad.onnx"):
-        target_dir = os.path.expanduser("~/.cache/whisper-live/")
-
-        # Ensure the target directory exists
-        os.makedirs(target_dir, exist_ok=True)
-
-        # Define the target file path
-        model_filename = os.path.join(target_dir, "silero_vad.onnx")
-
-        # Check if the model file already exists
-        if not os.path.exists(model_filename):
-            # If it doesn't exist, download the model using wget
-            print("Downloading VAD ONNX model...")
-            try:
-                subprocess.run(["wget", "-O", model_filename, model_url], check=True)
-            except subprocess.CalledProcessError:
-                print("Failed to download the model using wget.")
-        return model_filename
-