waxal-lin-mms-1b

MMS-1B native Lingala adapter, fine-tuned for Lingala as part of the WAXAL ASR solution.

Released artifact: top-3 uniform average of steps 1004 / 1506 / 1757. Not decoded in the shipped pipeline — this model embeds voices for TTIA.

Files

File	Bytes
model.safetensors	3,858,957,656
config.json	2,072
preprocessor_config.json	254
vocab.json	505
requirements.txt	—

Total weights: 3,858,960,487 bytes (3.86 GB).

Role in the pipeline

The TTIA voice embedder: the enrolment gallery and every test clip pass through its encoder, and mean+std-pooled hidden states (layers 4 and 8) drive the idiolect-profile match. It decodes no text in the shipped pipeline.

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

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

repo = "DariusTheGeek/waxal-lin-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	Lingala
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.