gustapoll/asr-paresi

MMS Paresi ASR (LoRA)

A LoRA adapter (r=8, alpha=16, targeting q_proj/v_proj) fine-tuning facebook/mms-1b-all for automatic speech recognition (ASR) in Paresi, a Brazilian indigenous language. Trained with a combined CTC + cycle-consistency (TTS resynthesis) + mel-reconstruction loss as part of the ablation variant no_mel.

Non-commercial use only

The base MMS model is licensed CC-BY-NC-4.0. This adapter inherits that restriction: non-commercial use only.

Quality caveat

This is an early-stage, low-resource fine-tune, evaluated on a small held-out set (n=30):

Metric	In-space	Out-of-space
CER	0.964	0.853
MCD	868.6	904.2

Treat outputs as a research demo, not production-quality transcription/synthesis.

Try it live

gustapoll/paresi-demo — a Gradio Space with both ASR and TTS tabs for Paresi, also callable as an API via gradio_client.

How to use

from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from peft import PeftModel

base = Wav2Vec2ForCTC.from_pretrained("facebook/mms-1b-all")
base.load_adapter('pab')  # MMS's own per-language adapter, load BEFORE the LoRA
model = PeftModel.from_pretrained(base, "REPO_ID").eval()
processor = Wav2Vec2Processor.from_pretrained("REPO_ID")