src/whisper_quant.py

import itertools
import logging
import os
import zlib

from typing import BinaryIO, Iterable, List, NamedTuple, Optional, Tuple, Union

import ctranslate2
import numpy as np
import tokenizers

from faster_whisper.audio import decode_audio
from faster_whisper.feature_extractor import FeatureExtractor
from faster_whisper.tokenizer import Tokenizer
from download_quantized import download_model, format_timestamp, get_logger
from faster_whisper.vad import (
    SpeechTimestampsMap,
    VadOptions,
    collect_chunks,
    get_speech_timestamps,
)


class Word(NamedTuple):
    start: float
    end: float
    word: str
    probability: float


class Segment(NamedTuple):
    id: int
    seek: int
    start: float
    end: float
    text: str
    tokens: List[int]
    temperature: float
    avg_logprob: float
    compression_ratio: float
    no_speech_prob: float
    words: Optional[List[Word]]


class TranscriptionOptions(NamedTuple):
    beam_size: int
    best_of: int
    patience: float
    length_penalty: float
    repetition_penalty: float
    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[Union[str, Iterable[int]]]
    prefix: Optional[str]
    suppress_blank: bool
    suppress_tokens: Optional[List[int]]
    without_timestamps: bool
    max_initial_timestamp: float
    word_timestamps: bool
    prepend_punctuations: str
    append_punctuations: str


class TranscriptionInfo(NamedTuple):
    language: str
    language_probability: float
    duration: float
    all_language_probs: Optional[List[Tuple[str, float]]]
    transcription_options: TranscriptionOptions
    vad_options: VadOptions


class WhisperModel:
    def __init__(
        self,
        model_size_or_path: str,
        device: str = "auto",
        device_index: Union[int, List[int]] = 0,
        compute_type: str = "default",
        cpu_threads: int = 0,
        num_workers: int = 1,
        download_root: Optional[str] = None,
        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), a path to a converted
            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.
            The model can also be loaded on multiple GPUs by passing a list of IDs
            (e.g. [0, 1, 2, 3]). In that case, multiple transcriptions can run in parallel
            when transcribe() is called from multiple Python threads (see also num_workers).
          compute_type: Type to use for computation.
            See https://opennmt.net/CTranslate2/quantization.html.
          cpu_threads: Number of threads to use when running on CPU (4 by default).
            A non zero value overrides the OMP_NUM_THREADS environment variable.
          num_workers: When transcribe() is called from multiple Python threads,
            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 models should be saved. If not set, the models
            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()

        if os.path.isdir(model_size_or_path):
            model_path = model_size_or_path
        else:
            model_path = download_model(
                model_size_or_path,
                local_files_only=local_files_only,
                cache_dir=download_root,
            )

        self.model = ctranslate2.models.Whisper(
            model_path,
            device=device,
            device_index=device_index,
            compute_type=compute_type,
            intra_threads=cpu_threads,
            inter_threads=num_workers,
        )

        tokenizer_file = os.path.join(model_path, "tokenizer.json")
        if 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.feature_extractor = FeatureExtractor()
        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
        )
        self.tokens_per_second = (
            self.feature_extractor.sampling_rate // self.num_samples_per_token
        )
        self.input_stride = 2
        self.time_precision = 0.02
        self.max_length = 448

    def transcribe(
        self,
        audio: Union[str, BinaryIO, np.ndarray],
        language: Optional[str] = None,
        task: str = "transcribe",
        beam_size: int = 5,
        best_of: int = 5,
        patience: float = 1,
        length_penalty: float = 1,
        repetition_penalty: float = 1,
        temperature: Union[float, List[float], Tuple[float, ...]] = [
            0.0,
            0.2,
            0.4,
            0.6,
            0.8,
            1.0,
        ],
        compression_ratio_threshold: Optional[float] = 2.4,
        log_prob_threshold: Optional[float] = -1.0,
        no_speech_threshold: Optional[float] = 0.6,
        condition_on_previous_text: bool = True,
        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],
        without_timestamps: bool = False,
        max_initial_timestamp: float = 1.0,
        word_timestamps: bool = False,
        prepend_punctuations: str = "\"'“¿([{-",
        append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
        vad_filter: bool = False,
        vad_parameters: Optional[Union[dict, VadOptions]] = None,
    ) -> Tuple[Iterable[Segment], TranscriptionInfo]:
        """Transcribes an input file.

        Arguments:
          audio: Path to the input file (or a file-like object), or the audio waveform.
          language: The language spoken in the audio. It should be a language code such
            as "en" or "fr". If not set, the language will be detected in the first 30 seconds
            of audio.
          task: Task to execute (transcribe or translate).
          beam_size: Beam size to use for decoding.
          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).
          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`.
          compression_ratio_threshold: If the gzip compression ratio is above this value,
            treat as failed.
          log_prob_threshold: If the average log probability over sampled tokens is
            below this value, treat as failed.
          no_speech_threshold: If the no_speech probability is higher than this value AND
            the average log probability over sampled tokens is below `log_prob_threshold`,
            consider the segment as silent.
          condition_on_previous_text: If True, the previous output of the model is provided
            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.
          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
            of symbols as defined in the model config.json file.
          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
            and dynamic time warping, and include the timestamps for each word in each segment.
          prepend_punctuations: If word_timestamps is True, merge these punctuation symbols
            with the next word
          append_punctuations: If word_timestamps is True, merge these punctuation symbols
            with the previous word
          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 or VadOptions class (see available
            parameters and default values in the class `VadOptions`).

        Returns:
          A tuple with:

            - a generator over transcribed segments
            - an instance of TranscriptionInfo
        """
        sampling_rate = self.feature_extractor.sampling_rate

        if not isinstance(audio, np.ndarray):
            audio = decode_audio(audio, sampling_rate=sampling_rate)

        duration = audio.shape[0] / sampling_rate

        self.logger.info(
            "Processing audio with duration %s", format_timestamp(duration)
        )

        if vad_filter:
            if vad_parameters is None:
                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)

            self.logger.info(
                "VAD filter removed %s of audio",
                format_timestamp(duration - (audio.shape[0] / sampling_rate)),
            )

            if self.logger.isEnabledFor(logging.DEBUG):
                self.logger.debug(
                    "VAD filter kept the following audio segments: %s",
                    ", ".join(
                        "[%s -> %s]"
                        % (
                            format_timestamp(chunk["start"] / sampling_rate),
                            format_timestamp(chunk["end"] / sampling_rate),
                        )
                        for chunk in speech_chunks
                    ),
                )

        else:
            speech_chunks = None

        features = self.feature_extractor(audio)

        encoder_output = None
        all_language_probs = None

        if language is None:
            if not self.model.is_multilingual:
                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]

                self.logger.info(
                    "Detected language '%s' with probability %.2f",
                    language,
                    language_probability,
                )
        else:
            language_probability = 1

        tokenizer = Tokenizer(
            self.hf_tokenizer,
            self.model.is_multilingual,
            task=task,
            language=language,
        )

        options = TranscriptionOptions(
            beam_size=beam_size,
            best_of=best_of,
            patience=patience,
            length_penalty=length_penalty,
            repetition_penalty=repetition_penalty,
            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]
            ),
            initial_prompt=initial_prompt,
            prefix=prefix,
            suppress_blank=suppress_blank,
            suppress_tokens=get_suppressed_tokens(tokenizer, suppress_tokens),
            without_timestamps=without_timestamps,
            max_initial_timestamp=max_initial_timestamp,
            word_timestamps=word_timestamps,
            prepend_punctuations=prepend_punctuations,
            append_punctuations=append_punctuations,
        )

        segments = self.generate_segments(features, tokenizer, options, encoder_output)

        if speech_chunks:
            segments = restore_speech_timestamps(segments, speech_chunks, sampling_rate)

        info = TranscriptionInfo(
            language=language,
            language_probability=language_probability,
            duration=duration,
            transcription_options=options,
            vad_options=vad_parameters,
            all_language_probs=all_language_probs,
        )

        return segments, info

    def generate_segments(
        self,
        features: np.ndarray,
        tokenizer: Tokenizer,
        options: TranscriptionOptions,
        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
        while seek < content_frames:
            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
            )
            segment_duration = segment_size * self.feature_extractor.time_per_frame

            if self.logger.isEnabledFor(logging.DEBUG):
                self.logger.debug(
                    "Processing segment at %s", format_timestamp(time_offset)
                )

            previous_tokens = all_tokens[prompt_reset_since:]
            prompt = self.get_prompt(
                tokenizer,
                previous_tokens,
                without_timestamps=options.without_timestamps,
                prefix=options.prefix if seek == 0 else None,
            )

            if encoder_output is None:
                encoder_output = self.encode(segment)

            (
                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
                should_skip = result.no_speech_prob > options.no_speech_threshold

                if (
                    options.log_prob_threshold is not None
                    and avg_logprob > options.log_prob_threshold
                ):
                    # don't skip if the logprob is high enough, despite the no_speech_prob
                    should_skip = False

                if should_skip:
                    self.logger.debug(
                        "No speech threshold is met (%f > %f)",
                        result.no_speech_prob,
                        options.no_speech_threshold,
                    )

                    # fast-forward to the next segment boundary
                    seek += segment_size
                    encoder_output = None
                    continue

            tokens = result.sequences_ids[0]

            previous_seek = seek
            current_segments = []

            single_timestamp_ending = (
                len(tokens) >= 2
                and tokens[-2] < tokenizer.timestamp_begin
                and tokens[-1] >= tokenizer.timestamp_begin
            )

            consecutive_timestamps = [
                i
                for i in range(len(tokens))
                if i > 0
                and tokens[i] >= tokenizer.timestamp_begin
                and tokens[i - 1] >= tokenizer.timestamp_begin
            ]

            if len(consecutive_timestamps) > 0:
                slices = list(consecutive_timestamps)
                if single_timestamp_ending:
                    slices.append(len(tokens))

                last_slice = 0
                for current_slice in slices:
                    sliced_tokens = tokens[last_slice:current_slice]
                    start_timestamp_position = (
                        sliced_tokens[0] - tokenizer.timestamp_begin
                    )
                    end_timestamp_position = (
                        sliced_tokens[-1] - tokenizer.timestamp_begin
                    )
                    start_time = (
                        time_offset + start_timestamp_position * self.time_precision
                    )
                    end_time = (
                        time_offset + end_timestamp_position * self.time_precision
                    )

                    current_segments.append(
                        dict(
                            seek=seek,
                            start=start_time,
                            end=end_time,
                            tokens=sliced_tokens,
                        )
                    )
                    last_slice = current_slice

                if single_timestamp_ending:
                    # single timestamp at the end means no speech after the last timestamp.
                    seek += segment_size
                else:
                    # otherwise, ignore the unfinished segment and seek to the last timestamp
                    last_timestamp_position = (
                        tokens[last_slice - 1] - tokenizer.timestamp_begin
                    )
                    seek += last_timestamp_position * self.input_stride

            else:
                duration = segment_duration
                timestamps = [
                    token for token in tokens if token >= tokenizer.timestamp_begin
                ]
                if len(timestamps) > 0 and timestamps[-1] != tokenizer.timestamp_begin:
                    last_timestamp_position = timestamps[-1] - tokenizer.timestamp_begin
                    duration = last_timestamp_position * self.time_precision

                current_segments.append(
                    dict(
                        seek=seek,
                        start=time_offset,
                        end=time_offset + duration,
                        tokens=tokens,
                    )
                )

                seek += segment_size

            if options.word_timestamps:
                self.add_word_timestamps(
                    current_segments,
                    tokenizer,
                    encoder_output,
                    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
                    )

                    if seek_shift > 0:
                        seek = previous_seek + seek_shift

            encoder_output = None

            for segment in current_segments:
                tokens = segment["tokens"]
                text = tokenizer.decode(tokens)

                if segment["start"] == segment["end"] or not text.strip():
                    continue

                all_tokens.extend(tokens)
                idx += 1

                yield 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
                    ),
                )

            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)

    def encode(self, features: np.ndarray) -> ctranslate2.StorageView:
        # 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.
        to_cpu = self.model.device == "cuda" and len(self.model.device_index) > 1

        features = np.expand_dims(features, 0)
        features = get_ctranslate2_storage(features)

        return self.model.encode(features, to_cpu=to_cpu)

    def generate_with_fallback(
        self,
        encoder_output: ctranslate2.StorageView,
        prompt: List[int],
        tokenizer: Tokenizer,
        options: TranscriptionOptions,
    ) -> Tuple[ctranslate2.models.WhisperGenerationResult, float, float, float]:
        decode_result = None
        all_results = []
        below_cr_threshold_results = []

        max_initial_timestamp_index = int(
            round(options.max_initial_timestamp / self.time_precision)
        )

        for temperature in options.temperatures:
            if temperature > 0:
                kwargs = {
                    "beam_size": 1,
                    "num_hypotheses": options.best_of,
                    "sampling_topk": 0,
                    "sampling_temperature": temperature,
                }
            else:
                kwargs = {
                    "beam_size": options.beam_size,
                    "patience": options.patience,
                }

            result = self.model.generate(
                encoder_output,
                [prompt],
                length_penalty=options.length_penalty,
                repetition_penalty=options.repetition_penalty,
                max_length=self.max_length,
                return_scores=True,
                return_no_speech_prob=True,
                suppress_blank=options.suppress_blank,
                suppress_tokens=options.suppress_tokens,
                max_initial_timestamp_index=max_initial_timestamp_index,
                **kwargs,
            )[0]

            tokens = result.sequences_ids[0]

            # Recover the average log prob from the returned score.
            seq_len = len(tokens)
            cum_logprob = result.scores[0] * (seq_len**options.length_penalty)
            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 options.compression_ratio_threshold is not None:
                if compression_ratio > options.compression_ratio_threshold:
                    needs_fallback = True  # too repetitive

                    self.logger.debug(
                        "Compression ratio threshold is not met with temperature %.1f (%f > %f)",
                        temperature,
                        compression_ratio,
                        options.compression_ratio_threshold,
                    )
                else:
                    below_cr_threshold_results.append(decode_result)

            if (
                options.log_prob_threshold is not None
                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_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 decode_result

    def get_prompt(
        self,
        tokenizer: Tokenizer,
        previous_tokens: List[int],
        without_timestamps: bool = False,
        prefix: Optional[str] = None,
    ) -> List[int]:
        prompt = []

        if previous_tokens:
            prompt.append(tokenizer.sot_prev)
            prompt.extend(previous_tokens[-(self.max_length // 2 - 1) :])

        prompt.extend(tokenizer.sot_sequence)

        if without_timestamps:
            prompt.append(tokenizer.no_timestamps)

        if prefix:
            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

    def add_word_timestamps(
        self,
        segments: List[dict],
        tokenizer: Tokenizer,
        encoder_output: ctranslate2.StorageView,
        num_frames: int,
        prepend_punctuations: str,
        append_punctuations: str,
        last_speech_timestamp: float,
    ):
        if len(segments) == 0:
            return

        text_tokens_per_segment = [
            [token for token in segment["tokens"] if token < tokenizer.eot]
            for segment in segments
        ]

        text_tokens = list(itertools.chain.from_iterable(text_tokens_per_segment))
        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 = (
            segments[0]["seek"]
            * self.feature_extractor.hop_length
            / self.feature_extractor.sampling_rate
        )

        word_index = 0

        for segment, text_tokens in zip(segments, text_tokens_per_segment):
            saved_tokens = 0
            words = []

            while word_index < len(alignment) and saved_tokens < len(text_tokens):
                timing = alignment[word_index]

                if timing["word"]:
                    words.append(
                        dict(
                            word=timing["word"],
                            start=round(time_offset + timing["start"], 2),
                            end=round(time_offset + timing["end"], 2),
                            probability=timing["probability"],
                        )
                    )

                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:
                # 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(
        self,
        tokenizer: Tokenizer,
        text_tokens: List[int],
        encoder_output: ctranslate2.StorageView,
        num_frames: int,
        median_filter_width: int = 7,
    ) -> List[dict]:
        if len(text_tokens) == 0:
            return []

        result = self.model.align(
            encoder_output,
            tokenizer.sot_sequence,
            [text_tokens],
            num_frames,
            median_filter_width=median_filter_width,
        )[0]

        text_token_probs = result.text_token_probs

        alignments = result.alignments
        text_indices = np.array([pair[0] for pair in alignments])
        time_indices = np.array([pair[1] for pair in alignments])

        words, word_tokens = tokenizer.split_to_word_tokens(
            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
        start_times = jump_times[word_boundaries[:-1]]
        end_times = jump_times[word_boundaries[1:]]
        word_probabilities = [
            np.mean(text_token_probs[i:j])
            for i, j in zip(word_boundaries[:-1], word_boundaries[1:])
        ]

        return [
            dict(
                word=word, tokens=tokens, start=start, end=end, probability=probability
            )
            for word, tokens, start, end, probability in zip(
                words, word_tokens, start_times, end_times, word_probabilities
            )
        ]


def restore_speech_timestamps(
    segments: Iterable[Segment],
    speech_chunks: List[dict],
    sampling_rate: int,
) -> Iterable[Segment]:
    ts_map = SpeechTimestampsMap(speech_chunks, sampling_rate)

    for segment in segments:
        if segment.words:
            words = []
            for word in segment.words:
                # Ensure the word start and end times are resolved to the same chunk.
                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),
                )
                words.append(word)

            segment = segment._replace(
                start=words[0].start,
                end=words[-1].end,
                words=words,
            )

        else:
            segment = segment._replace(
                start=ts_map.get_original_time(segment.start),
                end=ts_map.get_original_time(segment.end),
            )

        yield segment


def get_ctranslate2_storage(segment: np.ndarray) -> ctranslate2.StorageView:
    segment = np.ascontiguousarray(segment)
    segment = ctranslate2.StorageView.from_array(segment)
    return segment


def get_compression_ratio(text: str) -> float:
    text_bytes = text.encode("utf-8")
    return len(text_bytes) / len(zlib.compress(text_bytes))


def get_suppressed_tokens(tokenizer, suppress_tokens):
    if not suppress_tokens or -1 in suppress_tokens:
        return suppress_tokens

    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,
            tokenizer.translate,
            tokenizer.sot,
            tokenizer.sot_prev,
            tokenizer.sot_lm,
        ]
    )

    return sorted(set(suppress_tokens))


def merge_punctuations(alignment: List[dict], prepended: str, appended: str):
    # merge prepended punctuations
    i = len(alignment) - 2
    j = len(alignment) - 1
    while i >= 0:
        previous = alignment[i]
        following = alignment[j]
        if previous["word"].startswith(" ") and previous["word"].strip() in prepended:
            # prepend it to the following word
            following["word"] = previous["word"] + following["word"]
            following["tokens"] = previous["tokens"] + following["tokens"]
            previous["word"] = ""
            previous["tokens"] = []
        else:
            j = i
        i -= 1

    # merge appended punctuations
    i = 0
    j = 1
    while j < len(alignment):
        previous = alignment[i]
        following = alignment[j]
        if not previous["word"].endswith(" ") and following["word"] in appended:
            # append it to the previous word
            previous["word"] = previous["word"] + following["word"]
            previous["tokens"] = previous["tokens"] + following["tokens"]
            following["word"] = ""
            following["tokens"] = []
        else:
            i = j
        j += 1