dondo-nano-twi-ewe

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

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).

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

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.