app.py

import os
import time

import gradio as gr
import spaces
import torch
from loguru import logger
from pydub import AudioSegment
from transformers import pipeline

is_hf = os.getenv("SYSTEM") == "spaces"

generate_kwargs = {
    "language": "Japanese",
    "do_sample": False,
    "num_beams": 1,
    "no_repeat_ngram_size": 3,
}


model_dict = {
    "whisper-large-v2": "openai/whisper-large-v2",
    "whisper-large-v3": "openai/whisper-large-v3",
    "whisper-large-v3-turbo": "openai/whisper-large-v3-turbo",
    "kotoba-whisper-v1.0": "kotoba-tech/kotoba-whisper-v1.0",
    "kotoba-whisper-v2.0": "kotoba-tech/kotoba-whisper-v2.0",
    "galgame-whisper-wip": (
        "litagin/galgame-whisper-wip"
        if is_hf
        else "../whisper_finetune/galgame-whisper"
    ),
}

logger.info("Initializing pipelines...")
pipe_dict = {
    k: pipeline(
        "automatic-speech-recognition",
        model=v,
        device="cuda" if torch.cuda.is_available() or is_hf else "cpu",
    )
    for k, v in model_dict.items()
}
logger.success("Pipelines initialized!")


@spaces.GPU
def transcribe_common(audio: str, model: str) -> tuple[str, float]:
    logger.info(f"Transcribing {audio} with {model}")
    # Get duration of audio
    duration = AudioSegment.from_file(audio).duration_seconds
    if duration > 15:
        return "Audio too long, limit is 15 seconds", 0
    start_time = time.time()
    result = pipe_dict[model](audio, generate_kwargs=generate_kwargs)["text"]
    end_time = time.time()
    logger.success(f"Transcribed {audio} with {model} in {end_time - start_time:.2f}s")
    logger.success(f"Result:\n{result}")
    return result, end_time - start_time


def transcribe_large_v2(audio) -> tuple[str, float]:
    return transcribe_common(audio, "whisper-large-v2")


def transcribe_large_v3(audio) -> tuple[str, float]:
    return transcribe_common(audio, "whisper-large-v3")


def transcribe_large_v3_turbo(audio) -> tuple[str, float]:
    return transcribe_common(audio, "whisper-large-v3-turbo")


def transcribe_kotoba_v1(audio) -> tuple[str, float]:
    return transcribe_common(audio, "kotoba-whisper-v1.0")


def transcribe_kotoba_v2(audio) -> tuple[str, float]:
    return transcribe_common(audio, "kotoba-whisper-v2.0")


def transcribe_galgame_whisper(audio) -> tuple[str, float]:
    return transcribe_common(audio, "galgame-whisper-wip")


logger.info("Warm-up...")
transcribe_large_v3_turbo("test.wav")
logger.success("Warm-up done!")

initial_md = """
# Galgame-Whisper (WIP) Demo

- https://huggingface.co/litagin/galgame-whisper-wip
- 日本語のみ対応
- 比較できるように他モデルもついでに試せる
- 現在0.1エポックくらい
- 音声は15秒まで

pipeのハイパラ:
```python
generate_kwargs = {
    "language": "Japanese",
    "do_sample": False,
    "num_beams": 1,
    "no_repeat_ngram_size": 3,
}
```
"""

with gr.Blocks() as app:
    gr.Markdown(initial_md)
    audio = gr.Audio(type="filepath")
    with gr.Row():
        with gr.Column():
            gr.Markdown("### Galgame-Whisper (WIP)")
            button_galgame = gr.Button("Transcribe with Galgame-Whisper (WIP)")
            time_galgame = gr.Textbox(label="Time taken")
            output_galgame = gr.Textbox(label="Result")
    with gr.Row():
        with gr.Column():
            gr.Markdown("### Whisper-Large-V2")
            button_v2 = gr.Button("Transcribe with Whisper-Large-V2")
            time_v2 = gr.Textbox(label="Time taken")
            output_v2 = gr.Textbox(label="Result")
        with gr.Column():
            gr.Markdown("### Whisper-Large-V3")
            button_v3 = gr.Button("Transcribe with Whisper-Large-V3")
            time_v3 = gr.Textbox(label="Time taken")
            output_v3 = gr.Textbox(label="Result")
        with gr.Column():
            gr.Markdown("### Whisper-Large-V3-Turbo")
            button_v3_turbo = gr.Button("Transcribe with Whisper-Large-V3-Turbo")
            time_v3_turbo = gr.Textbox(label="Time taken")
            output_v3_turbo = gr.Textbox(label="Result")
    with gr.Row():
        with gr.Column():
            gr.Markdown("### Kotoba-Whisper-V1.0")
            button_kotoba_v1 = gr.Button("Transcribe with Kotoba-Whisper-V1.0")
            time_kotoba_v1 = gr.Textbox(label="Time taken")
            output_kotoba_v1 = gr.Textbox(label="Result")
        with gr.Column():
            gr.Markdown("### Kotoba-Whisper-V2.0")
            button_kotoba_v2 = gr.Button("Transcribe with Kotoba-Whisper-V2.0")
            time_kotoba_v2 = gr.Textbox(label="Time taken")
            output_kotoba_v2 = gr.Textbox(label="Result")

    with gr.Row():
        refresh_button = gr.Button("Refresh Status")  # Create a refresh button

    button_v2.click(transcribe_large_v2, inputs=audio, outputs=[output_v2, time_v2])
    button_v3.click(transcribe_large_v3, inputs=audio, outputs=[output_v3, time_v3])
    button_v3_turbo.click(
        transcribe_large_v3_turbo,
        inputs=audio,
        outputs=[output_v3_turbo, time_v3_turbo],
    )
    button_kotoba_v1.click(
        transcribe_kotoba_v1, inputs=audio, outputs=[output_kotoba_v1, time_kotoba_v1]
    )
    button_kotoba_v2.click(
        transcribe_kotoba_v2, inputs=audio, outputs=[output_kotoba_v2, time_kotoba_v2]
    )
    button_galgame.click(
        transcribe_galgame_whisper,
        inputs=audio,
        outputs=[output_galgame, time_galgame],
    )
app.launch(inbrowser=True)