sergyDwhiz/speachLTE

Lightweight speech-to-text model for Pidgin and Igbo, built to demonstrate accurate ASR for African low-resource languages that runs efficiently even on low-end devices. # SpeachLTE Cameroonian Pidgin + Ewondo speech-to-text that still runs on low-end devices. The repo ships a tiny Conformer-CTC model, the data prep it needs, and a demo that actually prints what the model hears. ## Why it exists Most Cameroonian audio—radio shows, town-hall recordings, WhatsApp notes—mixes Pidgin, French, and local languages like Ewondo. This project gives researchers and product teams a hackable, low-resource ASR stack so they can caption local content, build accessibility tools, or power voice commands without waiting on massive commercial APIs. ## What you actually get - **Training pipeline** – Hydra-driven script that builds manifests (synthetic by default), extracts mel features with torchaudio, and trains a compact Conformer. Output: `artifacts/base-conformer/best.ckpt`. - **Data plumbing** – JSONL manifests, text normalization, and tokenizer utilities ready for FLEURS/Common Voice or your own recordings. - **Demo output** – `scripts/demo_infer.py` loads any checkpoint, runs greedy decoding on a sample clip, and writes the transcript to `outputs/demo_transcript.txt` so you can see tangible words, not just logs. ## Run it (5‑minute tour) ```bash python -m venv .venv source .venv/bin/activate pip install -e .[dev] # 1. Train a quick model (uses synthetic audio if you haven't prepared real data yet) python3.10 scripts/train_model.py training.epochs=1 training.batch_size=2 model.num_layers=2 # 2. Show the transcript produced by that checkpoint python3.10 scripts/demo_infer.py --checkpoint artifacts/base-conformer/best.ckpt cat outputs/demo_transcript.txt ``` Need real data? Drop your curated audio + transcripts into `data/raw`, run `scripts/prepare_data.py`, and the same commands train on them. ### Use Common Voice Pidgin (real data) ```bash export HF_TOKEN=hf_xxx # Common Voice download requires a Hugging Face token python3.10 scripts/download_datasets.py --datasets common_voice --language pcm --cv-max-samples 200 python3.10 scripts/pre …