app.py

import whisper
import gradio as gr
from whisper_decoder_with_lm import LMOptions
from huggingface_hub import hf_hub_download

# Download model + LM from HF
model_path = hf_hub_download("sulaimank/whisper-small-lg-lm", "whisper-small-CV-Fleurs-lg-313hrs-v1.pt")
lm_path = hf_hub_download("sulaimank/whisper-small-lg-lm", "5gram.bin")

# Configure LM options
LMOptions().lm_path = lm_path
LMOptions().lm_alpha = 0.5  # update with best value
LMOptions().lm_beta = 0.5   # update with best value

# Load the model
model = whisper.load_model(model_path)

def transcribe(audio):
    decode_options = {
        "language": "sw",  # Swahili used as tokenizer workaround
        "without_timestamps": True,
        "temperature": 0.0,
        "beam_size": 5,
    }
    result = model.transcribe(audio, **decode_options)
    return result["text"]

interface = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(type="filepath", label="Upload Luganda Audio"),
    outputs="text",
    title="Luganda Whisper ASR with Language Model",
    description="Fine-tuned Whisper-small + 5-gram LM for Luganda transcription.",
)

interface.launch()