app.py

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import gradio as gr
from pydub import AudioSegment, silence
import tempfile
import torch
import torchaudio

MODEL_NAME = "islomov/navaistt_v2_medium"

processor = WhisperProcessor.from_pretrained(MODEL_NAME)
model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME)

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

def split_on_silence_with_duration_control(audio, min_len, max_len, silence_thresh=-40):
    silences = silence.detect_silence(audio, min_silence_len=500, silence_thresh=silence_thresh)
    silences = [((start + end) // 2) for start, end in silences]

    chunks = []
    start = 0
    while start < len(audio):
        end = min(start + max_len, len(audio))
        candidates = [s for s in silences if start + min_len <= s <= end]
        split_point = candidates[-1] if candidates else end
        chunks.append(audio[start:split_point])
        start = split_point
    return chunks

def transcribe(audio_file):
    # Load audio using pydub
    audio = AudioSegment.from_file(audio_file)
    
    # Convert to mono and 16kHz if needed
    if audio.channels > 1:
        audio = audio.set_channels(1)
    if audio.frame_rate != 16000:
        audio = audio.set_frame_rate(16000)

    # Detect silent chunks
    chunks = split_on_silence_with_duration_control(
        audio, min_len=15000, max_len=25000, silence_thresh=-40
    )

    # Transcribe each chunk
    results = []
    for chunk in chunks:
        with tempfile.NamedTemporaryFile(suffix=".wav", delete=True) as tmpfile:
            chunk.export(tmpfile.name, format="wav")
            waveform, _ = torchaudio.load(tmpfile.name)
            input_features = processor(
                waveform.squeeze().numpy(),
                sampling_rate=16000,
                return_tensors="pt",
                language="uz"
            ).input_features.to(device)

            with torch.no_grad():
                predicted_ids = model.generate(input_features)
                transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
                results.append(transcription)

    return " ".join(results)

demo = gr.Blocks()

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(type="filepath", label="Audio file"),
    outputs="text",
    title="Navai SST v2 Medium",
    description="Transcribe Uzbek speech to text using Navai SST v1 Medium model. Upload an audio file to get started. Model: https://huggingface.co/islomov/navaistt_v2_medium"
)

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

demo.launch()