app.py

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

MODEL_NAME = "dataprizma/whisper-large-v3-turbo"
BATCH_SIZE = 8

device = 0 if torch.cuda.is_available() else "cpu"

pipe = pipeline(
    task="automatic-speech-recognition",
    model=MODEL_NAME,
    chunk_length_s=9,
    device=device,
    model_kwargs={
        "attn_implementation": "eager"
    },
)

def transcribe(audio_file):
    if audio_file is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting.")

    with open(audio_file, "rb") as f:
        audio_data = f.read()

    audio_array = ffmpeg_read(audio_data, sampling_rate=pipe.feature_extractor.sampling_rate)
    duration = len(audio_array) / pipe.feature_extractor.sampling_rate
    print(f"Audio duration: {duration:.2f} seconds")

    result = pipe(
        inputs=audio_array,
        batch_size=BATCH_SIZE,
        return_timestamps=False,
        generate_kwargs={
            "task": "transcribe",
            "no_speech_threshold": 0.4,
            "logprob_threshold": -1.0,
            "compression_ratio_threshold": 2.4
        }
    )

    return result["text"] if isinstance(result, dict) else result

demo = gr.Blocks()

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(type="filepath", label="Audio file"),
    outputs="text",
    title="Whisper Large V3: Transcribe Audio",
    description="Whisper Large V3 fine-tuned for Uzbek language by Dataprizma",
)

with demo:
    gr.TabbedInterface([file_transcribe], ["Audio file"])

demo.launch()