app.py

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

mdl = "models/amithm3/whisper-medium"
processor = WhisperProcessor.from_pretrained(mdl, task="transcribe")
feature_extractor = WhisperFeatureExtractor.from_pretrained(mdl, task="transcribe")
model = WhisperForConditionalGeneration.from_pretrained(mdl)
sampling_rate = 16000


def transcribe(audio, language):
    audio, orig_freq = torchaudio.load(audio)
    audio = torchaudio.functional.resample(audio, orig_freq=orig_freq, new_freq=sampling_rate)
    audio = audio.squeeze().numpy()
    input_features = processor(audio, sampling_rate=sampling_rate, return_tensors="pt").input_features
    model.generation_config.language = language
    predicted_ids = model.generate(input_features)
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    return transcription


iface = gr.Interface(
    fn=transcribe,
    inputs=[gr.Audio(type="filepath"), gr.Dropdown(["kannada", "english", None], label="Language", value="kannada")],
    outputs="text",
    title="Whisper Medium Indic",
    description="Realtime demo for Indic speech recognition using a fine-tuned Whisper Medium model.",
)

iface.launch()