inference_example.py

from nemo.collections.asr.models import EncDecMultiTaskModel
from omegaconf import OmegaConf
import os
import argparse

def main():
    parser = argparse.ArgumentParser(description="Finnish ASR Inference Example")
    parser.add_argument("--audio", type=str, required=True, help="Path to the audio file (.wav)")
    parser.add_argument("--model", type=str, default="models/canary-finnish.nemo", help="Path to the finetuned .nemo model")
    parser.add_argument("--kenlm", type=str, default="models/kenlm_5M.nemo", help="Path to the KenLM model")
    parser.add_argument("--beam_size", type=int, default=4, help="Beam size for decoding")
    parser.add_argument("--pnc", type=str, default="yes", help="Enable Punctuation and Capitalization (yes/no)")
    
    args = parser.parse_args()

    # 1. Load Model and KenLM Bundle
    if not os.path.exists(args.model):
        print(f"Error: Model not found at {args.model}")
        return

    print(f"Loading model from {args.model}...")
    model = EncDecMultiTaskModel.restore_from(args.model)

    # Configure KenLM if provided
    if args.kenlm and os.path.exists(args.kenlm):
        print(f"Configuring decoding strategy with KenLM from {args.kenlm}...")
        model.change_decoding_strategy(
            decoding_cfg=OmegaConf.create({
                'strategy': 'beam',
                'beam': {
                    'beam_size': args.beam_size,
                    'ngram_lm_model': args.kenlm,
                    'ngram_lm_alpha': 0.2,
                },
                'batch_size': 1
            })
        )
    else:
        print("Using greedy decoding (no KenLM found or specified).")

    # 2. Transcribe with Finnish Prompts
    if not os.path.exists(args.audio):
        print(f"Error: Audio sample not found at {args.audio}")
        return

    print(f"Transcribing {args.audio}...")
    transcription = model.transcribe(
        audio=[args.audio],
        taskname="asr",
        source_lang="fi",
        target_lang="fi",
        pnc=args.pnc
    )

    print("-" * 30)
    print(f"Result: {transcription[0]}")
    print("-" * 30)

if __name__ == "__main__":
    main()