app.py

import gradio as gr

import torch
import librosa
import json
# Load model directly
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq

processor = AutoProcessor.from_pretrained("dmatekenya/whisper-large-v3-chichewa")
model = AutoModelForSpeechSeq2Seq.from_pretrained("dmatekenya/whisper-large-v3-chichewa")

def transcribe(audio_file_mic=None, audio_file_upload=None, language="English (eng)"):
    if audio_file_mic:
        audio_file = audio_file_mic
    elif audio_file_upload:
        audio_file = audio_file_upload
    else:
        return "Please upload an audio file or record one"

    # Make sure audio is 16kHz
    speech, sample_rate = librosa.load(audio_file)
    if sample_rate != 16000:
        speech = librosa.resample(speech, orig_sr=sample_rate, target_sr=16000)

    # Keep the same model in memory and simply switch out the language adapters by calling load_adapter() for the model and set_target_lang() for the tokenizer
    # language_code = iso_codes[language]
    # processor.tokenizer.set_target_lang(language_code)
    # model.load_adapter(language_code)

    inputs = processor(speech, sampling_rate=16_000, return_tensors="pt")

    with torch.no_grad():
        outputs = model(**inputs).logits

    ids = torch.argmax(outputs, dim=-1)[0]
    transcription = processor.decode(ids)
    return transcription


description = ''''''

iface = gr.Interface(fn=transcribe,
                     inputs=[
                         gr.Audio(source="microphone", type="filepath", label="Record Audio"),
                         gr.Audio(source="upload", type="filepath", label="Upload Audio"),
                         ],
                     outputs=gr.Textbox(label="Transcription"),
                     description=description
                     )
iface.launch()