gustapoll/asr-xavante

MMS Xavante 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 Xavante, a Brazilian indigenous language. Trained with a combined CTC + cycle-consistency (TTS resynthesis) + mel-reconstruction loss as part of the ablation variant full.

Note: Xavante is not in MMS's native language set. This adapter was bootstrapped from MMS's xer checkpoint (facebook/mms-1b-all) as the closest available base, then fine-tuned on Xavante data.

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=122):

Metric	In-space	Out-of-space
CER	0.524	0.523
MCD	938.6	931.3

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

Try it live

gustapoll/xavante-demo — a Gradio Space with both ASR and TTS tabs for Xavante, 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('xer')  # MMS's own per-language adapter, load BEFORE the LoRA
model = PeftModel.from_pretrained(base, "REPO_ID").eval()
processor = Wav2Vec2Processor.from_pretrained("REPO_ID")