app.py

import os
import tempfile
import time

import gradio as gr
import numpy as np
import torch
import yt_dlp as youtube_dl
from gradio_client import Client
from pyannote.audio import Pipeline
from transformers.pipelines.audio_utils import ffmpeg_read


YT_LENGTH_LIMIT_S = 36000  # limit to 1 hour YouTube files
SAMPLING_RATE = 16000

API_URL = "https://sanchit-gandhi-whisper-jax.hf.space/"
HF_TOKEN = os.environ.get("HF_TOKEN")

# set up the Gradio client
client = Client(API_URL)

# set up the diarization pipeline
diarization_pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization", use_auth_token=HF_TOKEN)


def format_string(timestamp):
    """
    Reformat a timestamp string from (HH:)MM:SS to float seconds. Note that the hour column
    is optional, and is appended within the function if not input.

    Args:
        timestamp (str):
            Timestamp in string format, either MM:SS or HH:MM:SS.
    Returns:
        seconds (float):
            Total seconds corresponding to the input timestamp.
    """
    split_time = timestamp.split(":")
    split_time = [float(sub_time) for sub_time in split_time]

    if len(split_time) == 2:
        split_time.insert(0, 0)

    seconds = split_time[0] * 3600 + split_time[1] * 60 + split_time[2]
    return seconds


# Adapted from https://github.com/openai/whisper/blob/c09a7ae299c4c34c5839a76380ae407e7d785914/whisper/utils.py#L50
def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = "."):
    """
    Reformat a timestamp from a float of seconds to a string in format (HH:)MM:SS. Note that the hour
    column is optional, and is appended in the function if the number of hours > 0.

    Args:
        seconds (float):
            Total seconds corresponding to the input timestamp.
    Returns:
        timestamp (str):
            Timestamp in string format, either MM:SS or HH:MM:SS.
    """
    if seconds is not None:
        milliseconds = round(seconds * 1000.0)

        hours = milliseconds // 3_600_000
        milliseconds -= hours * 3_600_000

        minutes = milliseconds // 60_000
        milliseconds -= minutes * 60_000

        seconds = milliseconds // 1_000
        milliseconds -= seconds * 1_000

        hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
        return f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
    else:
        # we have a malformed timestamp so just return it as is
        return seconds


def format_as_transcription(raw_segments):
    return "\n\n".join(
        [
            f"{chunk['speaker']} [{format_timestamp(chunk['timestamp'][0])} -> {format_timestamp(chunk['timestamp'][1])}] {chunk['text']}"
            for chunk in raw_segments
        ]
    )


def _return_yt_html_embed(yt_url):
    video_id = yt_url.split("?v=")[-1]
    HTML_str = (
        f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
        " </center>"
    )
    return HTML_str


def download_yt_audio(yt_url, filename):
    info_loader = youtube_dl.YoutubeDL()
    try:
        info = info_loader.extract_info(yt_url, download=False)
    except youtube_dl.utils.DownloadError as err:
        raise gr.Error(str(err))

    file_length = info["duration_string"]
    file_length_s = format_string(file_length)

    if file_length_s > YT_LENGTH_LIMIT_S:
        yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))
        file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))
        raise gr.Error(f"To encourage fair usage of the demo, the maximum YouTube length is {yt_length_limit_hms}, "
                       f"got {file_length_hms} YouTube video.")

    ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
    with youtube_dl.YoutubeDL(ydl_opts) as ydl:
        try:
            ydl.download([yt_url])
        except youtube_dl.utils.ExtractorError as err:
            raise gr.Error(str(err))


def align(transcription, segments, group_by_speaker=True):
    transcription_split = transcription.split("\n")

    # re-format transcription from string to List[Dict]
    transcript = []
    for chunk in transcription_split:
        start_end, transcription = chunk[1:].split("] ")
        start, end = start_end.split("->")

        transcript.append({"timestamp": (format_string(start), format_string(end)), "text": transcription})

    # diarizer output may contain consecutive segments from the same speaker (e.g. {(0 -> 1, speaker_1), (1 -> 1.5, speaker_1), ...})
    # we combine these segments to give overall timestamps for each speaker's turn (e.g. {(0 -> 1.5, speaker_1), ...})
    new_segments = []
    prev_segment = cur_segment = segments[0]

    for i in range(1, len(segments)):
        cur_segment = segments[i]

        # check if we have changed speaker ("label")
        if cur_segment["label"] != prev_segment["label"] and i < len(segments):
            # add the start/end times for the super-segment to the new list
            new_segments.append(
                {
                    "segment": {"start": prev_segment["segment"]["start"], "end": cur_segment["segment"]["start"]},
                    "speaker": prev_segment["label"],
                }
            )
            prev_segment = segments[i]

    # add the last segment(s) if there was no speaker change
    new_segments.append(
        {
            "segment": {"start": prev_segment["segment"]["start"], "end": cur_segment["segment"]["end"]},
            "speaker": prev_segment["label"],
        }
    )

    # get the end timestamps for each chunk from the ASR output
    end_timestamps = np.array([chunk["timestamp"][-1] for chunk in transcript])
    segmented_preds = []

    # align the diarizer timestamps and the ASR timestamps
    for segment in new_segments:
        # get the diarizer end timestamp
        end_time = segment["segment"]["end"]
        # find the ASR end timestamp that is closest to the diarizer's end timestamp and cut the transcript to here
        upto_idx = np.argmin(np.abs(end_timestamps - end_time))

        if group_by_speaker:
            segmented_preds.append(
                {
                    "speaker": segment["speaker"],
                    "text": "".join([chunk["text"] for chunk in transcript[: upto_idx + 1]]),
                    "timestamp": (transcript[0]["timestamp"][0], transcript[upto_idx]["timestamp"][1]),
                }
            )
        else:
            for i in range(upto_idx + 1):
                segmented_preds.append({"speaker": segment["speaker"], **transcript[i]})

        # crop the transcripts and timestamp lists according to the latest timestamp (for faster argmin)
        transcript = transcript[upto_idx + 1 :]
        end_timestamps = end_timestamps[upto_idx + 1 :]

    # final post-processing
    transcription = format_as_transcription(segmented_preds)
    return transcription


def transcribe(audio_path, task="transcribe", group_by_speaker=True, progress=gr.Progress()):
    # run Whisper JAX asynchronously using Gradio client (endpoint)
    job = client.submit(
        audio_path,
        task,
        True,
        api_name="/predict_1",
    )

    # run diarization while we wait for Whisper JAX
    diarization = diarization_pipeline(audio_path)
    segments = diarization.for_json()["content"]

    # only fetch the transcription result after performing diarization
    transcription, _ = job.result()

    # align the ASR transcriptions and diarization timestamps
    transcription = align(transcription, segments, group_by_speaker=group_by_speaker)

    return transcription


def transcribe_yt(yt_url, task="transcribe", group_by_speaker=True, progress=gr.Progress()):
    # run Whisper JAX asynchronously using Gradio client (endpoint)
    job = client.submit(
        yt_url,
        task,
        True,
        api_name="/predict_2",
    )

    html_embed_str = _return_yt_html_embed(yt_url)
    with tempfile.TemporaryDirectory() as tmpdirname:
        filepath = os.path.join(tmpdirname, "video.mp4")
        download_yt_audio(yt_url, filepath)
        with open(filepath, "rb") as f:
            inputs = f.read()

    inputs = ffmpeg_read(inputs, SAMPLING_RATE)
    inputs = torch.from_numpy(inputs).float()
    inputs = inputs.unsqueeze(0)

    diarization = diarization_pipeline(
        {"waveform": inputs, "sample_rate": SAMPLING_RATE},
    )
    segments = diarization.for_json()["content"]

    # only fetch the transcription result after performing diarization
    _, transcription, _ = job.result()

    # align the ASR transcriptions and diarization timestamps
    transcription = align(transcription, segments, group_by_speaker=group_by_speaker)

    return html_embed_str, transcription


title = "Whisper JAX + Speaker Diarization ⚡️"

description = """Combine the speed of Whisper JAX with pyannote speaker diarization to transcribe meetings in super fast time. Demo uses Whisper JAX as an [endpoint](https://twitter.com/sanchitgandhi99/status/1656665496463495168) and pyannote speaker diarization running locally. The Whisper JAX endpoint is run asynchronously, meaning speaker diarization is run in parallel to the speech transcription. The diarized timestamps are aligned with the Whisper output to give the final speaker-segmented transcription.

To duplicate the demo, first accept the pyannote terms of use for the [speaker diarization](https://huggingface.co/pyannote/speaker-diarization) and [segmentation](https://huggingface.co/pyannote/segmentation) models. Then, click [here](https://huggingface.co/spaces/sanchit-gandhi/whisper-jax-diarization?duplicate=true) to duplicate the demo, and enter your Hugging Face access token as a Space secret when prompted.
"""

article = "Whisper large-v2 model by OpenAI. Speaker diarization model by pyannote. Whisper JAX backend running JAX on a TPU v4-8 through the generous support of the [TRC](https://sites.research.google/trc/about/) programme. Whisper JAX [code](https://github.com/sanchit-gandhi/whisper-jax) and Gradio demo by 🤗 Hugging Face."

microphone = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.inputs.Audio(source="microphone", optional=True, type="filepath"),
        gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
        gr.inputs.Checkbox(default=True, label="Group by speaker"),
    ],
    outputs=[
        gr.outputs.Textbox(label="Transcription").style(show_copy_button=True),
    ],
    allow_flagging="never",
    title=title,
    description=description,
    article=article,
)

audio_file = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.inputs.Audio(source="upload", optional=True, label="Audio file", type="filepath"),
        gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
        gr.inputs.Checkbox(default=True, label="Group by speaker"),
    ],
    outputs=[
        gr.outputs.Textbox(label="Transcription").style(show_copy_button=True),
    ],
    allow_flagging="never",
    title=title,
    description=description,
    article=article,
)

youtube = gr.Interface(
    fn=transcribe_yt,
    inputs=[
        gr.inputs.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
        gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
        gr.inputs.Checkbox(default=True, label="Group by speaker"),
    ],
    outputs=[
        gr.outputs.HTML(label="Video"),
        gr.outputs.Textbox(label="Transcription").style(show_copy_button=True),
    ],
    allow_flagging="never",
    title=title,
    examples=[["https://www.youtube.com/watch?v=m8u-18Q0s7I", "transcribe", True]],
    cache_examples=False,
    description=description,
    article=article,
)

demo = gr.Blocks()

with demo:
    gr.TabbedInterface([microphone, audio_file, youtube], ["Microphone", "Audio File", "YouTube"])

demo.queue(max_size=10)
demo.launch()