billingsmoore/garchen-stt

Model Card for Model ID

Speech-to-Text model for Garchen Rinpoche.

Usage

The wav2vec and associated kenlm can be loaded like this:

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from pyctcdecode import build_ctcdecoder
from huggingface_hub import hf_hub_download
import json
import torch

# Load processor and model
processor = Wav2Vec2Processor.from_pretrained("billingsmoore/garchen-stt")
model = Wav2Vec2ForCTC.from_pretrained("billingsmoore/garchen-stt")

# Download KenLM model and decoder config from Hugging Face Hub
kenlm_path = hf_hub_download(repo_id="billingsmoore/garchen-stt", filename="kenlm_5gram.bin")
decoder_config_path = hf_hub_download(repo_id="billingsmoore/garchen-stt", filename="decoder_config.json")

# Load decoder config
with open(decoder_config_path, 'r') as f:
    decoder_config = json.load(f)

# Build decoder
vocab = processor.tokenizer.get_vocab()
labels = [token for token, idx in sorted(vocab.items(), key=lambda x: x[1])]

decoder = build_ctcdecoder(
    labels=labels,
    kenlm_model_path=kenlm_path,
    alpha=decoder_config["alpha"],
    beta=decoder_config["beta"]
)

Then, assuming that your data has some batch['audio'] = = {"array": waveform, "sampling_rate": sr}, predictions can be generated like this:

def transcribe_with_lm(audio, processor=processor, model=model, decoder=decoder, decoder_config=decoder_config):
        # Extract features
        inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
        
        # Get model predictions
        with torch.no_grad():
            logits = model(inputs.input_values).logits.cpu().numpy()[0]
        
        # Decode with LM (hyperparameters from grid search)
        transcription = decoder.decode(logits,
                                        beam_width=decoder_config["beam_width"],
                                        beam_prune_logp=decoder_config["beam_prune_logp"],
                                        token_min_logp=decoder_config["token_min_logp"])
        return transcription

transcriptions = [transcribe_with_lm(batch["audio"]["array"]) for batch in dataset]