DONDO-nanos: on-device ASR for Twi and Ewe

Name: dondo-nano-twi-ewe
Availability: InStock

Compressed, on-device-friendly versions of Khaya AI's DONDO speech model for Twi and Ewe. The default model in this repo is Nano-L12 (315.6M params). Smaller variants and an int8 ONNX build are in subfolders.

The recipe is depth reduction with a warm start: a shallower Wav2Vec2-BERT student is initialized from evenly-spaced layers of the 24-layer parent (plus its feature projection, adapter, and CTC head), then fine-tuned on WAXAL Twi and Ewe. Knowledge distillation from the parent was tested and dropped, because the parent is out-of-domain on WAXAL and its soft targets made the student worse.

Results

Blended error = 0.5·WER + 0.5·CER on WAXAL validation (Twi/Ewe mean), greedy CTC decode, normalized text. All rows below are trained on the same WAXAL data (CTC, no distillation).

Model	Layers	Params	Disk (fp32)	CPU RTF	Blended
Fine-tuned parent	24	605.8M	2423 MB	0.252	0.216
Nano-L12 (default)	12	315.6M	1262 MB	0.132	0.261
Nano-L6 (`nano-l6/`)	6	170.5M	682 MB	0.069	0.387
Nano-L3 (`nano-l3/`)	3	98.0M	392 MB	0.036	~0.57

Halving the model (606M to 316M) costs 0.045 blended error. Nano-L12 also beats a zero-shot Meta MMS-1b (~965M, ~0.37 on these clips), though that comparison favors the nano, which saw in-domain data.

int8 (onnx/): ONNX Runtime dynamic int8 shrinks Nano-L12 from 1262 MB to 319 MB (~4x) for almost no accuracy change (0.266 to 0.269 in the same runtime). It does not speed up inference on CPU, since the model is already faster than real time; the win is size.

Usage

python

from transformers import AutoProcessor, AutoModelForCTC
import torch, librosa

proc = AutoProcessor.from_pretrained("samwell/dondo-nano-twi-ewe")          # Nano-L12
model = AutoModelForCTC.from_pretrained("samwell/dondo-nano-twi-ewe").eval()
# smaller variants: subfolder="nano-l6" or "nano-l3"

y, _ = librosa.load("clip.wav", sr=16000)
feats = proc(y, sampling_rate=16000, return_tensors="pt").input_features
with torch.no_grad():
    ids = model(feats).logits.argmax(-1)
print(proc.batch_decode(ids)[0])

Limitations

Evaluated on WAXAL validation only; no on-device (ARM) latency numbers yet.
Trained on ~6k clips; more data would likely close the Nano-L12 to parent gap (0.261 vs 0.216).
Twi and Ewe only, from a parent that also covers other Ghanaian languages.

Attribution

Base model: KhayaAI/w2v-bert-ada_ewe_fat_fra_gaa_nzi_twi_en (Apache-2.0)
Training data: google/WaxalNLP (CC-BY-SA-4.0 / CC-BY-4.0)

Released under CC-BY-SA-4.0 to respect the share-alike terms of the WAXAL training data.