app.py

import gradio as gr
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torchaudio

def translate(audio):
    model_id_asr = "openai/whisper-small"
    processor_asr = WhisperProcessor.from_pretrained(model_id_asr)
    model_asr = WhisperForConditionalGeneration.from_pretrained(model_id_asr)
    forced_decoder_ids = processor_asr.get_decoder_prompt_ids(language="tamil", task="translate")
    input_features = processor_asr(audio["audio"]["array"], sampling_rate=audio["audio"]["sampling_rate"], return_tensors="pt").input_features
    predicted_ids = model_asr.generate(input_features,forced_decoder_ids=forced_decoder_ids)
    transcription = processor_asr.batch_decode(predicted_ids, skip_special_tokens=True)
    return transcription[0]

def speech_to_speech_translation(audio_filepath):
    waveform, sampling_rate = torchaudio.load(audio_filepath)
    if sampling_rate != 16000:
      resampler = torchaudio.transforms.Resample(orig_freq=sampling_rate, new_freq=16000)
      waveform = resampler(waveform)
      sampling_rate = 16000
    audio_dict = {
        "audio": {
            "array": waveform.numpy(),
            "sampling_rate": sampling_rate
        }
    }
    translated_text = translate(audio_dict)
    return translated_text

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="microphone", type="filepath"),
    outputs="text",allow_flagging="never")

with demo:
    gr.TabbedInterface([mic_translate], ["Local Tamil Translator"])
demo.launch(debug=True, share=False)