kiviki/romani-asr-experiments

Romani ASR Experiments

This repository collects the reproducible training and evaluation artifacts for the Romani ASR experiments.

It does not contain raw audio, full transcript manifests, or private training data. Model weights live in model-specific repositories.

Current Model Repositories

• Whisper Turbo Romani LoRA adapter: kiviki/whisper-turbo-romani-lora
• MMS adapter: not published yet. The current MMS result is zero-shot facebook/mms-1b-all with target_lang=rmc-script_latin; a real MMS adapter still needs GPU training.

Naming Convention

• Prompt Test = test.csv, the original single-source prompt set (R_prompts). Use it for continuity with the original Whisper model card.
• Multi-Recording Dev = validation_clean.csv, the cleaned multi-source validation split. Use it for iteration and model selection.
• Whisper Turbo Baseline = openai/whisper-large-v3-turbo with a Slovak transcription prompt and no Romani fine-tuning.
• Whisper Turbo Romani LoRA = the public fine-tuned Whisper adapter.
• MMS-1B-All Romani/Latin Zero-Shot = public facebook/mms-1b-all with target_lang=rmc-script_latin, with no local MMS training yet.

Frozen Evaluation

All scores below were rerun with fixed manifests and guarded decoding on 2026-08-10. Lower WER/CER is better.

Split	System	Role	WER	CER	ASCII WER	ASCII CER	Non-Latin
Prompt Test	Whisper Turbo Baseline	unadapted baseline	0.796	0.184	0.719	0.138	0
Prompt Test	Whisper Turbo Romani LoRA	trained Whisper adapter	0.653	0.122	0.606	0.101	0
Prompt Test	MMS-1B-All Romani/Latin Zero-Shot	external zero-shot comparison	0.300	0.047	0.293	0.046	0
Multi-Recording Dev	Whisper Turbo Baseline	unadapted baseline	0.921	0.496	0.882	0.459	0
Multi-Recording Dev	Whisper Turbo Romani LoRA	trained Whisper adapter	0.769	0.405	0.753	0.393	0
Multi-Recording Dev	MMS-1B-All Romani/Latin Zero-Shot	external zero-shot comparison	0.756	0.383	0.746	0.376	0

Prompt Test contains only one source group, so it is useful for continuity with the original Whisper model card but too easy to be the only headline result. Multi-Recording Dev spans multiple recording groups and is the better iteration target.

Detailed report:

• reports/frozen-asr-evaluation-2026-08-10.md

Machine-readable metrics:

• eval/frozen-2026-08-10/*/metrics.json

GPU Training

The Mac can run MMS evaluation and smoke tests, but real MMS-1B adapter training should run on CUDA because the Wav2Vec2 CTC loss falls back to CPU on Apple Silicon.

Start here:

• docs/mms-gpu-training.md
• notebooks/mms_gpu_training_colab.ipynb
• scripts/run_mms_gpu_training.sh
• scripts/train_mms_adapter.py

The intended MMS output is a small adapter file:

adapter.rmc-script_latin.safetensors
processor/
training_summary.json

Do not upload or redistribute the full facebook/mms-1b-all base model in this project. Users should load the public base model and then the small adapter.

Repository Contents

docs/        GPU training guide
eval/        frozen metrics JSON files
notebooks/   Colab workflow
reports/     human-readable evaluation reports
scripts/     train/evaluate/package/summarize scripts
src/         helper modules used by the scripts
tests/       lightweight unit tests for local helpers

Raw audio, full manifests, prediction CSVs, and the local GPU ZIP bundle are intentionally excluded.