VoiceBridge — TORGO three-speaker adapter

This is the shipped VoiceBridge LoRA adapter: an adaptation of Aaliya05/voicebridge on the three TORGO speakers used in the demo (M02, M04, M05).

• Base model: openai/whisper-small
• Adapter: PEFT LoRA, r=16, alpha=32, q_proj + v_proj
• Size: ~7.1 MB
• Training: continued from Aaliya05/voicebridge on 443 utterances from M02/M04/M05.
• Excluded: all 11 rows whose reference text matches one of the three demo sentences, so the demo clips and their transcripts were not trained on.

Scope — read this

This is speaker-dependent adaptation, not a general dysarthria model. It was trained and evaluated on the same three TORGO speakers; no fourth speaker has been evaluated. Severe speech from an unseen speaker remains unsupported. Live microphone results vary with channel and room.

Measured on the three demo clips (with the product's benchmark):

Clip	WER
Mild (M05)	0.0%
Moderate-to-severe (M04)	16.7%
Severe (M02)	0.0%

Mean clip WER 5.6%; corpus WER 6.2%. See the GitHub repository's LIMITS.md and facts.json for the full caveats.

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

base = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
model = PeftModel.from_pretrained(base, "Nomad1331/voicebridge")
processor = WhisperProcessor.from_pretrained("openai/whisper-small")

Licensing

The adapter is derived from TORGO recordings and Aaliya05/voicebridge; the TORGO dataset is distributed under the Linguistic Data Consortium user agreement. Check redistribution terms before publishing a fork. The base model is MIT-licensed.