divy318/multilingual-stt-model

🎙️ Proprietary Multilingual Speech-to-Text Model (Gujarati • Hindi • English)

A production-grade trilingual acoustic model fine-tuned from Meta MMS-300M (Wav2Vec2ForCTC) on 41.77 hours of verified trilingual speech audio.

📊 Performance Benchmark

Language	Clean CER	Clean WER	Audio Hours
Gujarati (gu)	13.16%	83.33%	16.90 Hours
Hindi (hi)	30.43%	140.0%	11.32 Hours
English (en)	13.16%	62.50%	13.55 Hours
Combined Corpus Record	10.30%	34.12%	41.77 Hours

🚀 Usage with Hugging Face Transformers

import torch
import soundfile as sf
from transformers import Wav2Vec2ForCTC

# Load model
model = Wav2Vec2ForCTC.from_pretrained("divy318/multilingual-stt-model")
model.eval()

# Load audio (16,000 Hz Mono)
waveform, sr = sf.read("sample.wav", dtype="float32")
inputs = torch.tensor(waveform).unsqueeze(0)

# Inference
with torch.no_grad():
    logits = model(inputs).logits
pred_ids = torch.argmax(logits, dim=-1)

📜 License

MIT License