transcribe.py

import argparse
import os
import warnings
from typing import List, Optional, Tuple, Union, Iterator, TYPE_CHECKING

import numpy as np
import torch
import torchaudio
from transformers import AutoProcessor, Wav2Vec2ForCTC
import tqdm
from .audio import SAMPLE_RATE, N_FRAMES, HOP_LENGTH, pad_or_trim, log_mel_spectrogram, load_audio
from .alignment import get_trellis, backtrack, merge_repeats, merge_words
from .decoding import DecodingOptions, DecodingResult
from .tokenizer import LANGUAGES, TO_LANGUAGE_CODE, get_tokenizer
from .utils import exact_div, format_timestamp, optional_int, optional_float, str2bool, write_txt, write_vtt, write_srt, write_ass

if TYPE_CHECKING:
    from .model import Whisper

LANGUAGES_WITHOUT_SPACES = ["ja", "zh"]

DEFAULT_ALIGN_MODELS_TORCH = {
    "en": "WAV2VEC2_ASR_BASE_960H",
    "fr": "VOXPOPULI_ASR_BASE_10K_FR",
    "de": "VOXPOPULI_ASR_BASE_10K_DE",
    "es": "VOXPOPULI_ASR_BASE_10K_ES",
    "it": "VOXPOPULI_ASR_BASE_10K_IT",
}

DEFAULT_ALIGN_MODELS_HF = {
    "ja": "jonatasgrosman/wav2vec2-large-xlsr-53-japanese",
}


def transcribe(
    model: "Whisper",
    audio: Union[str, np.ndarray, torch.Tensor],
    *,
    verbose: Optional[bool] = None,
    temperature: Union[float, Tuple[float, ...]] = (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
    compression_ratio_threshold: Optional[float] = 2.4,
    logprob_threshold: Optional[float] = -1.0,
    no_speech_threshold: Optional[float] = 0.6,
    condition_on_previous_text: bool = False, # turn off by default due to errors it causes
    **decode_options,
):
    """
    Transcribe an audio file using Whisper

    Parameters
    ----------
    model: Whisper
        The Whisper model instance

    audio: Union[str, np.ndarray, torch.Tensor]
        The path to the audio file to open, or the audio waveform

    verbose: bool
        Whether to display the text being decoded to the console. If True, displays all the details,
        If False, displays minimal details. If None, does not display anything

    temperature: Union[float, Tuple[float, ...]]
        Temperature for sampling. It can be a tuple of temperatures, which will be successfully used
        upon failures according to either `compression_ratio_threshold` or `logprob_threshold`.

    compression_ratio_threshold: float
        If the gzip compression ratio is above this value, treat as failed

    logprob_threshold: float
        If the average log probability over sampled tokens is below this value, treat as failed

    no_speech_threshold: float
        If the no_speech probability is higher than this value AND the average log probability
        over sampled tokens is below `logprob_threshold`, consider the segment as silent

    condition_on_previous_text: bool
        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.

    decode_options: dict
        Keyword arguments to construct `DecodingOptions` instances

    Returns
    -------
    A dictionary containing the resulting text ("text") and segment-level details ("segments"), and
    the spoken language ("language"), which is detected when `decode_options["language"]` is None.
    """
    dtype = torch.float16 if decode_options.get("fp16", True) else torch.float32
    if model.device == torch.device("cpu"):
        if torch.cuda.is_available():
            warnings.warn("Performing inference on CPU when CUDA is available")
        if dtype == torch.float16:
            warnings.warn("FP16 is not supported on CPU; using FP32 instead")
            dtype = torch.float32

    if dtype == torch.float32:
        decode_options["fp16"] = False

    mel = log_mel_spectrogram(audio)

    if decode_options.get("language", None) is None:
        if not model.is_multilingual:
            decode_options["language"] = "en"
        else:
            if verbose:
                print("Detecting language using up to the first 30 seconds. Use `--language` to specify the language")
            segment = pad_or_trim(mel, N_FRAMES).to(model.device).to(dtype)
            _, probs = model.detect_language(segment)
            decode_options["language"] = max(probs, key=probs.get)
            if verbose is not None:
                print(f"Detected language: {LANGUAGES[decode_options['language']].title()}")

    language = decode_options["language"]
    task = decode_options.get("task", "transcribe")
    tokenizer = get_tokenizer(model.is_multilingual, language=language, task=task)

    def decode_with_fallback(segment: torch.Tensor) -> DecodingResult:
        temperatures = [temperature] if isinstance(temperature, (int, float)) else temperature
        decode_result = None

        for t in temperatures:
            kwargs = {**decode_options}
            if t > 0:
                # disable beam_size and patience when t > 0
                kwargs.pop("beam_size", None)
                kwargs.pop("patience", None)
            else:
                # disable best_of when t == 0
                kwargs.pop("best_of", None)

            options = DecodingOptions(**kwargs, temperature=t)
            decode_result = model.decode(segment, options)

            needs_fallback = False
            if compression_ratio_threshold is not None and decode_result.compression_ratio > compression_ratio_threshold:
                needs_fallback = True  # too repetitive
            if logprob_threshold is not None and decode_result.avg_logprob < logprob_threshold:
                needs_fallback = True  # average log probability is too low

            if not needs_fallback:
                break

        return decode_result

    seek = 0
    input_stride = exact_div(
        N_FRAMES, model.dims.n_audio_ctx
    )  # mel frames per output token: 2
    time_precision = (
        input_stride * HOP_LENGTH / SAMPLE_RATE
    )  # time per output token: 0.02 (seconds)
    all_tokens = []
    all_segments = []
    prompt_reset_since = 0

    initial_prompt = decode_options.pop("initial_prompt", None) or []
    if initial_prompt:
        initial_prompt = tokenizer.encode(" " + initial_prompt.strip())
        all_tokens.extend(initial_prompt)

    def add_segment(
        *, start: float, end: float, text_tokens: torch.Tensor, result: DecodingResult
    ):
        text = tokenizer.decode([token for token in text_tokens if token < tokenizer.eot])
        if len(text.strip()) == 0:  # skip empty text output
            return

        all_segments.append(
            {
                "id": len(all_segments),
                "seek": seek,
                "start": start,
                "end": end,
                "text": text,
                "tokens": text_tokens.tolist(),
                "temperature": result.temperature,
                "avg_logprob": result.avg_logprob,
                "compression_ratio": result.compression_ratio,
                "no_speech_prob": result.no_speech_prob,
            }
        )
        if verbose:
            print(f"[{format_timestamp(start)} --> {format_timestamp(end)}] {text}")

    # show the progress bar when verbose is False (otherwise the transcribed text will be printed)
    num_frames = mel.shape[-1]
    previous_seek_value = seek

    with tqdm.tqdm(total=num_frames, unit='frames', disable=verbose is not False) as pbar:
        while seek < num_frames:
            timestamp_offset = float(seek * HOP_LENGTH / SAMPLE_RATE)
            segment = pad_or_trim(mel[:, seek:], N_FRAMES).to(model.device).to(dtype)
            segment_duration = segment.shape[-1] * HOP_LENGTH / SAMPLE_RATE

            decode_options["prompt"] = all_tokens[prompt_reset_since:]
            result: DecodingResult = decode_with_fallback(segment)
            tokens = torch.tensor(result.tokens)

            if no_speech_threshold is not None:
                # no voice activity check
                should_skip = result.no_speech_prob > no_speech_threshold
                if logprob_threshold is not None and result.avg_logprob > logprob_threshold:
                    # don't skip if the logprob is high enough, despite the no_speech_prob
                    should_skip = False

                if should_skip:
                    seek += segment.shape[-1]  # fast-forward to the next segment boundary
                    continue

            timestamp_tokens: torch.Tensor = tokens.ge(tokenizer.timestamp_begin)
            consecutive = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0].add_(1)
            if len(consecutive) > 0:  # if the output contains two consecutive timestamp tokens
                last_slice = 0
                for current_slice in consecutive:
                    sliced_tokens = tokens[last_slice:current_slice]
                    start_timestamp_position = (
                        sliced_tokens[0].item() - tokenizer.timestamp_begin
                    )
                    end_timestamp_position = (
                        sliced_tokens[-1].item() - tokenizer.timestamp_begin
                    )
                    add_segment(
                        start=timestamp_offset + start_timestamp_position * time_precision,
                        end=timestamp_offset + end_timestamp_position * time_precision,
                        text_tokens=sliced_tokens[1:-1],
                        result=result,
                    )
                    last_slice = current_slice
                last_timestamp_position = (
                    tokens[last_slice - 1].item() - tokenizer.timestamp_begin
                )
                seek += last_timestamp_position * input_stride
                all_tokens.extend(tokens[: last_slice + 1].tolist())
            else:
                duration = segment_duration
                timestamps = tokens[timestamp_tokens.nonzero().flatten()]
                if len(timestamps) > 0 and timestamps[-1].item() != tokenizer.timestamp_begin:
                    # no consecutive timestamps but it has a timestamp; use the last one.
                    # single timestamp at the end means no speech after the last timestamp.
                    last_timestamp_position = timestamps[-1].item() - tokenizer.timestamp_begin
                    duration = last_timestamp_position * time_precision

                add_segment(
                    start=timestamp_offset,
                    end=timestamp_offset + duration,
                    text_tokens=tokens,
                    result=result,
                )

                seek += segment.shape[-1]
                all_tokens.extend(tokens.tolist())

            if not condition_on_previous_text or result.temperature > 0.5:
                # do not feed the prompt tokens if a high temperature was used
                prompt_reset_since = len(all_tokens)

            # update progress bar
            pbar.update(min(num_frames, seek) - previous_seek_value)
            previous_seek_value = seek

    return dict(text=tokenizer.decode(all_tokens[len(initial_prompt):]), segments=all_segments, language=language)


def align(
    transcript: Iterator[dict],
    language: str,
    model: torch.nn.Module,
    align_model_metadata: dict,
    audio: Union[str, np.ndarray, torch.Tensor],
    device: str,
    extend_duration: float = 0.0,
    start_from_previous: bool = True,
    drop_non_aligned_words: bool = False,
):
    print("Performing alignment...")
    if not torch.is_tensor(audio):
        if isinstance(audio, str):
            audio = load_audio(audio)
        audio = torch.from_numpy(audio)
    if len(audio.shape) == 1:
        audio = audio.unsqueeze(0)

    MAX_DURATION = audio.shape[1] / SAMPLE_RATE

    model_dictionary = align_model_metadata['dictionary']
    model_lang = align_model_metadata['language']
    model_type = align_model_metadata['type']

    prev_t2 = 0
    word_segments_list = []
    for idx, segment in enumerate(transcript):
        t1 = max(segment['start'] - extend_duration, 0)
        t2 = min(segment['end'] + extend_duration, MAX_DURATION)
        if start_from_previous and t1 < prev_t2:
            t1 = prev_t2

        f1 = int(t1 * SAMPLE_RATE)
        f2 = int(t2 * SAMPLE_RATE)

        waveform_segment = audio[:, f1:f2]
        with torch.inference_mode():
            if model_type == "torchaudio":
                emissions, _ = model(waveform_segment.to(device))
            elif model_type == "huggingface":
                emissions = model(waveform_segment.to(device)).logits
            else:
                raise NotImplementedError(f"Align model of type {model_type} not supported.")
            emissions = torch.log_softmax(emissions, dim=-1)
        emission = emissions[0].cpu().detach()
        transcription = segment['text'].strip()
        if language not in LANGUAGES_WITHOUT_SPACES:
            t_words = transcription.split(' ')
        else:
            t_words = [c for c in transcription]

        t_words_clean = [''.join([w for w in word if w.lower() in model_dictionary.keys()]) for word in t_words]
        t_words_nonempty = [x for x in t_words_clean if x != ""]
        t_words_nonempty_idx = [x for x in range(len(t_words_clean)) if t_words_clean[x] != ""]
        segment['word-level'] = []

        if len(t_words_nonempty) > 0:
            transcription_cleaned = "|".join(t_words_nonempty).lower()
            tokens = [model_dictionary[c] for c in transcription_cleaned]
            trellis = get_trellis(emission, tokens)
            path = backtrack(trellis, emission, tokens)
            segments = merge_repeats(path, transcription_cleaned)
            word_segments = merge_words(segments)
            ratio = waveform_segment.size(0) / (trellis.size(0) - 1)

            duration = t2 - t1
            local = []
            t_local = [None] * len(t_words)
            for wdx, word in enumerate(word_segments):
                t1_ = ratio * word.start
                t2_ = ratio * word.end
                local.append((t1_, t2_))
                t_local[t_words_nonempty_idx[wdx]] = (t1_ * duration + t1, t2_ * duration + t1)
            t1_actual = t1 + local[0][0] * duration
            t2_actual = t1 + local[-1][1] * duration

            segment['start'] = t1_actual
            segment['end'] = t2_actual
            prev_t2 = segment['end']

            # for the .ass output
            for x in range(len(t_local)):
                curr_word = t_words[x]
                curr_timestamp = t_local[x]
                if curr_timestamp is not None:
                    segment['word-level'].append({"text": curr_word, "start": curr_timestamp[0], "end": curr_timestamp[1]})
                else:
                    segment['word-level'].append({"text": curr_word, "start": None, "end": None})

            # for per-word .srt ouput
            # merge missing words to previous, or merge with next word ahead if idx == 0
            for x in range(len(t_local)):
                curr_word = t_words[x]
                curr_timestamp = t_local[x]
                if curr_timestamp is not None:
                    word_segments_list.append({"text": curr_word, "start": curr_timestamp[0], "end": curr_timestamp[1]})
                elif not drop_non_aligned_words:
                    # then we merge
                    if x == 0:
                        t_words[x+1] = " ".join([curr_word, t_words[x+1]])
                    else:
                        word_segments_list[-1]['text'] += ' ' + curr_word
        else:
            # then we resort back to original whisper timestamps
            # segment['start] and segment['end'] are unchanged
            prev_t2 = 0
            segment['word-level'].append({"text": segment['text'], "start": segment['start'], "end":segment['end']})
            word_segments_list.append({"text": segment['text'], "start": segment['start'], "end":segment['end']})

        print(f"[{format_timestamp(segment['start'])} --> {format_timestamp(segment['end'])}] {segment['text']}")

    return {"segments": transcript, "word_segments": word_segments_list}

def load_align_model(language_code, device, model_name=None):
    if model_name is None:
        # use default model
        if language_code in DEFAULT_ALIGN_MODELS_TORCH:
            model_name = DEFAULT_ALIGN_MODELS_TORCH[language_code]
        elif language_code in DEFAULT_ALIGN_MODELS_HF:
            model_name = DEFAULT_ALIGN_MODELS_HF[language_code]
        else:
            print(f"There is no default alignment model set for this language ({language_code}).\
                Please find a wav2vec2.0 model finetuned on this language in https://huggingface.co/models, then pass the model name in --align_model [MODEL_NAME]")
            raise ValueError(f"No default align-model for language: {language_code}")

    if model_name in torchaudio.pipelines.__all__:
        pipeline_type = "torchaudio"
        bundle = torchaudio.pipelines.__dict__[model_name]
        align_model = bundle.get_model().to(device)
        labels = bundle.get_labels()
        align_dictionary = {c.lower(): i for i, c in enumerate(labels)}
    else:
        try:
            processor = AutoProcessor.from_pretrained(model_name)
            align_model = Wav2Vec2ForCTC.from_pretrained(model_name)
        except Exception as e:
            print(e)
            print(f"Error loading model from huggingface, check https://huggingface.co/models for finetuned wav2vec2.0 models")
            raise ValueError(f'The chosen align_model "{model_name}" could not be found in huggingface (https://huggingface.co/models) or torchaudio (https://pytorch.org/audio/stable/pipelines.html#id14)')
        pipeline_type = "huggingface"
        align_model = align_model.to(device)
        labels = processor.tokenizer.get_vocab()
        align_dictionary = {char.lower(): code for char,code in processor.tokenizer.get_vocab().items()}

    align_metadata = {"language": language_code, "dictionary": align_dictionary, "type": pipeline_type}

    return align_model, align_metadata

def cli():
    from . import available_models

    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument("audio", nargs="+", type=str, help="audio file(s) to transcribe")
    parser.add_argument("--model", default="small", choices=available_models(), help="name of the Whisper model to use")
    parser.add_argument("--model_dir", type=str, default=None, help="the path to save model files; uses ~/.cache/whisper by default")
    parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu", help="device to use for PyTorch inference")
    # alignment params
    parser.add_argument("--align_model", default=None, help="Name of phoneme-level ASR model to do alignment")
    parser.add_argument("--align_extend", default=2, type=float, help="Seconds before and after to extend the whisper segments for alignment")
    parser.add_argument("--align_from_prev", default=True, type=bool, help="Whether to clip the alignment start time of current segment to the end time of the last aligned word of the previous segment")
    parser.add_argument("--drop_non_aligned", action="store_true", help="For word .srt, whether to drop non aliged words, or merge them into neighbouring.")

    parser.add_argument("--output_dir", "-o", type=str, default=".", help="directory to save the outputs")
    parser.add_argument("--output_type", default="srt", choices=['all', 'srt', 'vtt', 'txt'], help="directory to save the outputs")

    parser.add_argument("--verbose", type=str2bool, default=True, help="whether to print out the progress and debug messages")

    parser.add_argument("--task", type=str, default="transcribe", choices=["transcribe", "translate"], help="whether to perform X->X speech recognition ('transcribe') or X->English translation ('translate')")
    parser.add_argument("--language", type=str, default=None, choices=sorted(LANGUAGES.keys()) + sorted([k.title() for k in TO_LANGUAGE_CODE.keys()]), help="language spoken in the audio, specify None to perform language detection")

    parser.add_argument("--temperature", type=float, default=0, help="temperature to use for sampling")
    parser.add_argument("--best_of", type=optional_int, default=5, help="number of candidates when sampling with non-zero temperature")
    parser.add_argument("--beam_size", type=optional_int, default=5, help="number of beams in beam search, only applicable when temperature is zero")
    parser.add_argument("--patience", type=float, default=None, help="optional patience value to use in beam decoding, as in https://arxiv.org/abs/2204.05424, the default (1.0) is equivalent to conventional beam search")
    parser.add_argument("--length_penalty", type=float, default=None, help="optional token length penalty coefficient (alpha) as in https://arxiv.org/abs/1609.08144, uses simple length normalization by default")

    parser.add_argument("--suppress_tokens", type=str, default="-1", help="comma-separated list of token ids to suppress during sampling; '-1' will suppress most special characters except common punctuations")
    parser.add_argument("--initial_prompt", type=str, default=None, help="optional text to provide as a prompt for the first window.")
    parser.add_argument("--condition_on_previous_text", type=str2bool, default=False, help="if True, provide the previous output of the model 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")
    parser.add_argument("--fp16", type=str2bool, default=True, help="whether to perform inference in fp16; True by default")

    parser.add_argument("--temperature_increment_on_fallback", type=optional_float, default=0.2, help="temperature to increase when falling back when the decoding fails to meet either of the thresholds below")
    parser.add_argument("--compression_ratio_threshold", type=optional_float, default=2.4, help="if the gzip compression ratio is higher than this value, treat the decoding as failed")
    parser.add_argument("--logprob_threshold", type=optional_float, default=-1.0, help="if the average log probability is lower than this value, treat the decoding as failed")
    parser.add_argument("--no_speech_threshold", type=optional_float, default=0.6, help="if the probability of the <|nospeech|> token is higher than this value AND the decoding has failed due to `logprob_threshold`, consider the segment as silence")
    parser.add_argument("--threads", type=optional_int, default=0, help="number of threads used by torch for CPU inference; supercedes MKL_NUM_THREADS/OMP_NUM_THREADS")

    args = parser.parse_args().__dict__
    model_name: str = args.pop("model")
    model_dir: str = args.pop("model_dir")
    output_dir: str = args.pop("output_dir")
    output_type: str = args.pop("output_type")
    device: str = args.pop("device")

    align_model: str = args.pop("align_model")
    align_extend: float = args.pop("align_extend")
    align_from_prev: bool = args.pop("align_from_prev")
    drop_non_aligned: bool = args.pop("drop_non_aligned")

    os.makedirs(output_dir, exist_ok=True)

    if model_name.endswith(".en") and args["language"] not in {"en", "English"}:
        if args["language"] is not None:
            warnings.warn(f"{model_name} is an English-only model but receipted '{args['language']}'; using English instead.")
        args["language"] = "en"

    temperature = args.pop("temperature")
    temperature_increment_on_fallback = args.pop("temperature_increment_on_fallback")
    if temperature_increment_on_fallback is not None:
        temperature = tuple(np.arange(temperature, 1.0 + 1e-6, temperature_increment_on_fallback))
    else:
        temperature = [temperature]

    threads = args.pop("threads")
    if threads > 0:
        torch.set_num_threads(threads)

    from . import load_model
    model = load_model(model_name, device=device, download_root=model_dir)

    align_language = args["language"] if args["language"] is not None else "en" # default to loading english if not specified
    align_model, align_metadata = load_align_model(align_language, device, model_name=align_model)

    for audio_path in args.pop("audio"):
        result = transcribe(model, audio_path, temperature=temperature, **args)

        if result["language"] != align_metadata["language"]:
            # load new language
            print(f"New language found ({result['language']})! Previous was ({align_metadata['language']}), loading new alignment model for new language...")
            align_model, align_metadata = load_align_model(result["language"], device)

        result_aligned = align(result["segments"], result["language"], align_model, align_metadata, audio_path, device,
                                extend_duration=align_extend, start_from_previous=align_from_prev, drop_non_aligned_words=drop_non_aligned)
        audio_basename = os.path.basename(audio_path)

        # save TXT
        if output_type in ["txt", "all"]:
            with open(os.path.join(output_dir, audio_basename + ".txt"), "w", encoding="utf-8") as txt:
                write_txt(result_aligned["segments"], file=txt)

        # save VTT
        if output_type in ["vtt", "all"]:
            with open(os.path.join(output_dir, audio_basename + ".vtt"), "w", encoding="utf-8") as vtt:
                write_vtt(result_aligned["segments"], file=vtt)

        # save SRT
        if output_type in ["srt", "all"]:
            with open(os.path.join(output_dir, audio_basename + ".srt"), "w", encoding="utf-8") as srt:
                write_srt(result_aligned["segments"], file=srt)

        # save per-word SRT
        with open(os.path.join(output_dir, audio_basename + ".word.srt"), "w", encoding="utf-8") as srt:
            write_srt(result_aligned["word_segments"], file=srt)

        # save ASS
        with open(os.path.join(output_dir, audio_basename + ".ass"), "w", encoding="utf-8") as srt:
            write_ass(result_aligned["segments"], file=srt)


if __name__ == '__main__':
    cli()