Member of the 4-family word-medoid MBR on the Shona lane; the medoid is chosen by distance to peers, so a member can shape the selection geometry without being selected itself.

This model is one component of an ensemble solution and is not intended to be used alone. The routing, decoding, fusion and post-processing pipeline are in the solution repository.

How to use

python

# pip install -r requirements.txt   (pinned; see Files)
import librosa, torch
from transformers import AutoProcessor, Wav2Vec2ForCTC

repo = "DariusTheGeek/waxal-sna-mms-1b"
processor = AutoProcessor.from_pretrained(repo)
model = Wav2Vec2ForCTC.from_pretrained(repo).eval().cuda()

audio, _ = librosa.load("clip.wav", sr=16_000, mono=True)
inputs = processor(audio, sampling_rate=16_000, return_tensors="pt")
with torch.no_grad():
    logits = model(inputs.input_values.cuda()).logits
print(processor.batch_decode(logits.argmax(-1))[0])

Provenance


Parent	`facebook/mms-1b-all`
Language	Shona
Fine-tuning data	Waxal Lingala/Shona supervised split (`google/WaxalNLP`)
Seed	42

Non-commercial. This licence is inherited from facebook/mms-1b-all and is binding on anyone who downloads these weights. The OmniASR models in this solution are apache-2.0; only the two MMS models carry the NC restriction.

Usage

See https://github.com/DariusTheGeek/waxal-asr-solution for the environment locks, decode configuration and the exact command that reproduces the submission end to end from audio.

Licence

cc-by-nc-4.0, inherited from the parent model.