app.py

import os
from math import floor
from typing import Optional

import spaces
import torch
import gradio as gr
from transformers import pipeline
from transformers.pipelines.audio_utils import ffmpeg_read


# configuration
MODEL_NAME = "kotoba-tech/kotoba-whisper-bilingual-v1.0"
BATCH_SIZE = 16
CHUNK_LENGTH_S = 15
# device setting
if torch.cuda.is_available():
    torch_dtype = torch.bfloat16
    device = "cuda"
    model_kwargs = {'attn_implementation': 'sdpa'}
else:
    torch_dtype = torch.float32
    device = "cpu"
    model_kwargs = {}

# define the pipeline
pipe = pipeline(
    model=MODEL_NAME,
    chunk_length_s=CHUNK_LENGTH_S,
    batch_size=BATCH_SIZE,
    torch_dtype=torch_dtype,
    device=device,
    model_kwargs=model_kwargs,
    trust_remote_code=True
)


def format_time(start: Optional[float], end: Optional[float]):

    def _format_time(seconds: Optional[float]):
        if seconds is None:
            return "complete    "
        minutes = floor(seconds / 60)
        hours = floor(seconds / 3600)
        seconds = seconds - hours * 3600 - minutes * 60
        m_seconds = floor(round(seconds - floor(seconds), 3) * 10 ** 3)
        seconds = floor(seconds)
        return f'{hours:02}:{minutes:02}:{seconds:02}.{m_seconds:03}'

    return f"[{_format_time(start)}-> {_format_time(end)}]:"


@spaces.GPU
def get_prediction(inputs, task: str, language: Optional[str]):
    generate_kwargs = {"task": task, "language": language}
    prediction = pipe(inputs, return_timestamps=True, generate_kwargs=generate_kwargs)
    text = "".join([c['text'] for c in prediction['chunks']])
    text_timestamped = "\n".join([f"{format_time(*c['timestamp'])} {c['text']}" for c in prediction['chunks']])
    return text, text_timestamped


def transcribe(inputs: str, task: str, language: str):
    if inputs is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
    with open(inputs, "rb") as f:
        inputs = f.read()
    inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
    return get_prediction(inputs, task, language)


demo = gr.Blocks()
description = (f"Kotoba-whisper-bilingual is speech transcribe and translation model for English and Japanese! Demo "
               f"[{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe/translate audio"
               f" files of arbitrary length. Make sure to choose the desired language of the transcription from the tab.")
title = f"Transcribe/translate Japanese & English Audio with Kotoba-Whisper-Bilingual"
mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources="microphone", type="filepath"),
        gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
        gr.Radio(["ja", "en"], label="Output Language", value="ja")
    ],
    outputs=["text", "text"],
    title=title,
    description=description,
    allow_flagging="never",
)
file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources="upload", type="filepath", label="Audio file"),
        gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
        gr.Radio(["ja", "en"], label="Output Language", value="ja")
    ],
    outputs=["text", "text"],
    title=title,
    description=description,
    allow_flagging="never",
)
with demo:
    gr.TabbedInterface([mf_transcribe, file_transcribe], ["Microphone", "Audio file"])
demo.queue(api_open=False, default_concurrency_limit=40).launch(show_api=False, show_error=True)