RobotsMali/quartznum-v0

QuartzNum v0

QuartzNum is an 18M-parameter QuartzNet 15x5 CTC model specialized for recognizing Bambara spoken amounts, account numbers, and phone numbers. It is fine-tuned from RobotsMali/stt-bm-quartznet15x5-v2 and is intended as a narrow component, not a replacement for general-purpose Bambara ASR.

The model is deployed in the mobilebamspeech integration demo and the voice-first mobileBAMking proof of concept. NeMoOnnxSharp provides the ONNX Runtime/.NET inference path used on Android.

Intended Use

Use QuartzNum to transcribe short, isolated number expressions in the vocabulary and speaking conditions represented by the banking voice collection. Pass its text through a Bambara number parser/normalizer and validate the resulting value before application use.

Do not use this research checkpoint by itself to authorize transactions, identify speakers, or transcribe unrestricted speech.

Use with NVIDIA NeMo

pip install "nemo-toolkit[asr]"

from nemo.collections.asr.models import EncDecCTCModel

model = EncDecCTCModel.from_pretrained("RobotsMali/quartznum-v0")
model.eval()
predictions = model.transcribe(["spoken_number.wav"])
print(predictions[0].text)

Input audio is converted by the model preprocessor to mono 16 kHz features. Output is a greedy CTC hypothesis over a lowercase character vocabulary containing the Bambara letters ŋ, ɔ, ɛ, and ɲ.

NeMo 2.5.0 Compatibility

This checkpoint was created with NeMo 2.5.0. Loading it with newer NeMo versions (observed with 2.7.x) can fail because the strict decoding schema expects key_phrase_items_list; see NVIDIA-NeMo/Speech#15658. The following workaround was tested with Python 3.12 and allows the 2.5.0 checkpoint to load in the newer environment:

from pathlib import Path

from nemo.collections.asr.models import ASRModel
from omegaconf import OmegaConf

model_name = "RobotsMali/quartznum-v0"
cfg = ASRModel.from_pretrained(model_name, return_config=True)
OmegaConf.set_struct(cfg, False)

for decoder in ("greedy", "beam"):
    boosting_tree = OmegaConf.select(cfg, f"decoding.{decoder}.boosting_tree")
    if boosting_tree is not None:
        boosting_tree.key_phrase_items_list = None

config_path = Path("patched_config.yaml").resolve()
OmegaConf.save(cfg, config_path)
model = ASRModel.from_pretrained(
    model_name=model_name,
    override_config_path=str(config_path),
    strict=False,
)
config_path.unlink()

Architecture and Training

QuartzNet 15x5 uses one-dimensional time-channel separable convolutions and a character CTC decoder. Training used NeMo 2.5.0, BF16 mixed precision, batches of 16, Adam at 5e-4, and cosine annealing with 80 warm-up steps. The run allowed up to 50 epochs with early stopping patience of 15. The training script, configuration, and archived result are public.

Training clips were derived from the 588-command MobileBAMking voice collection. For transfer commands, the preparation pipeline extracts separately recorded amount and account/phone-number audio and normalizes the labels with bambara-normalizer. This application-collected training set is not currently published as a standalone Hugging Face dataset.

Evaluation

The archived held-out evaluation reports:

Split	Decoding	WER (%) ↓	Test loss
MobileBAMking spoken-number test split	Greedy CTC	17.24	36.29

WER measures token transcription, not end-to-end numeric-value accuracy. The split is small, narrow-domain, and drawn from the same collection workflow as training, so the figure should not be generalized to arbitrary speakers or acoustic conditions.

Limitations

• Long digit strings can be costly even when only one character is wrong; validate parsed values explicitly.
• Accuracy may degrade with noise, overlap, dialects, recording hardware, or number forms absent from training.
• The model does not provide confidence calibration, speaker verification, or transaction safety controls.
• General Bambara words and sentences are outside its intended domain.

License

The model is released under CC BY 4.0.