PrinceAlhassanNasamu/tekyerema-asr-mms-kus

Kusaal ASR (MMS adapter)

Author: Prince Nasamu Alhassan

Overview

A per-language MMS adapter over facebook/mms-1b-all.

This is now the best Kusaal recogniser in the project. The previous champion, KASA-42, scores 41.0 / 19.2 and cannot be fine-tuned from — it ships raw .pt files and an inference-only int8 ONNX export with no architectures or model_type. This model beats it by 10.6 WER points and can be trained further.

Use it

This repo holds only the adapter (about 9 MB), not a whole model. Load the base and apply it — and take the head width from the adapter, not from the tokenizer: they differ by the added special tokens, and letting the tokenizer decide raises a size mismatch.

AutoConfig.from_pretrained on this repo fails with "Unrecognized model … should have a model_type key", and AutoModelForCTC fails too. That is expected, not a broken upload: there is no model here, only an adapter and its tokenizer. AutoProcessor.from_pretrained does work. A config.json is deliberately not shipped — the base's would advertise a vocabulary of 154 where this head is 58–62 wide, and would invite loading weights that are not in the repo.

import torch, soundfile as sf
from huggingface_hub import hf_hub_download, list_repo_files
from safetensors.torch import load_file
from transformers import AutoProcessor, Wav2Vec2ForCTC

repo = "PrinceAlhassanNasamu/tekyerema-asr-mms-kus"
adapter = [f for f in list_repo_files(repo)
           if f.startswith("adapter.") and f.endswith(".safetensors")][0]
sd = load_file(hf_hub_download(repo, adapter))

proc  = AutoProcessor.from_pretrained(repo)
model = Wav2Vec2ForCTC.from_pretrained(
    "facebook/mms-1b-all",
    vocab_size=sd["lm_head.weight"].shape[0],   # the ADAPTER decides this
    ignore_mismatched_sizes=True).eval()

missing, unexpected = model.load_state_dict(sd, strict=False)
assert not unexpected, unexpected   # never transcribe with a half-loaded model

wav, sr = sf.read("clip.wav", dtype="float32")
inp = proc(wav, sampling_rate=16_000, return_tensors="pt")
with torch.no_grad():
    logits = model(**inp).logits
print(proc.batch_decode(logits.argmax(-1))[0])

Training data

Trained on the Ghana Speech dataset and related Ghanaian corpora, licensed CC BY-NC 4.0.

Measured

On kusaal_scripture, same items and same scorer as the baseline:

model	WER / CER
baseline it was fine-tuned from	59.34 / 22.10
this model	30.44 / 13.52

A WER alone is not informative — compared against another language it means nothing. Compared against the model it started from, it means everything.

Intended use & license

Non-commercial use only (CC BY-NC 4.0). This is inherited from the training data and required by the terms under which the compute was granted: models trained in that window are non-commercial by condition of access, not by inference.

Limitations, stated plainly

• Dagbani had no recogniser of its own for this whole project, and the reason given for that was wrong. Every card here said "one fine-tuning session on 74 validation rows would not change that". Those 74 rows are the eng-dag machine-translation validation split. The Dagbani speech data in this same account is waxal_dag: 13,228 training rows, 1,750 validation rows, ~71 hours, 1,041 speakers with the largest at 1% — more data and better speaker diversity than Ewe, which produced a working 42.19 WER recogniser. A number was carried across from a translation table into a speech claim, and then repeated on every model card on the account. It is training now, on 2026-08-31. Until it is scored, the honest statement is that Dagbani's best available recogniser scores 86.6 WER and nobody had tried fine-tuning on the data already in hand.
• Evaluation is on read and machine-translated text. No recordings of people speaking agent commands in these languages exist. Numbers measured this way are optimistic about phrasing and pessimistic about code-switching, and should not be read as field performance.
• Research work from a hackathon entry, not a supported product.

The rest of the family

Recognisers

• whisper-large-v3-turbo-tekyerema-eng-foundation — Ghanaian English ASR — course 1 (foundation)
• kusaal-whisper-small-lora — Kusaal ASR (Whisper-small LoRA, superseded)
• kasa42-asr — KASA-42 (Kusaal, third-party export)
• tekyerema-asr-ctc — Twi ASR (w2v-BERT CTC)
• tekyerema-asr-mms-ewe — Ewe ASR (MMS adapter)
• tekyerema-asr-mms-dag — Dagbani ASR (MMS adapter)
• tekyerema-asr-mms-hau — Hausa ASR (MMS adapter)
• tekyerema-asr-mms-kus — Kusaal ASR (MMS adapter)
• whisper-large-v3-turbo-tekyerema-eng — Ghanaian English ASR (Whisper large-v3-turbo)

Voices

• tekyerema-tts-twi — Twi TTS (VITS)
• tekyerema-tts-kus — Kusaal TTS (VITS)
• tekyerema-tts-ewe — Ewe TTS (VITS)
• tekyerema-tts-hau — Hausa TTS (VITS)
• tekyerema-tts-eng — Ghanaian English TTS (VITS)

Agent models

• tekyerema-1-reply — Tɛkyerɛma-1 reply adapter (arm ①)
• tekyerema-1-native-reply — Tɛkyerɛma-1 reply adapter (arm ②)
• tekyerema-1-tool — Tɛkyerɛma-1 tool adapter (arm 1)
• tekyerema-audio-native — Tɛkyerɛma-1 audio-native (arm 3)
• tekyerema-audio-native-4k — Tɛkyerɛma-1 audio-native, 4,000 clips (arm 3 v2)
• tekyerema-1-native-tool — Tɛkyerɛma-1 tool adapter (arm 2)

Translation

• tekyerema-nllb600m-v1 — Tɛkyerɛma MT v1 (NLLB-600M)
• kusaal-nllb-600M — Kusaal MT specialist (NLLB-600M)

Routing

• tekyerema-intent-afroxlmr — Intent classifier (AfroXLMR)

Acknowledgements

Compute resources provided by AI Skills and Compute Africa (AISCA). Trained on the Ghana NLP H200 GPU. Please keep derivatives non-commercial and share improvements back with the Ghana NLP community (ghananlpcommunity).