Griot Nano 1
Griot Nano 1 is a 153M-parameter multilingual Conformer CTC model for Akan, Dagbani, Ewe, Ghanaian English, and Ga. Metrics below use raw greedy CTC decoding without using the suggested KenLM.
Test results
| Language | Samples | CER | WER |
|---|---|---|---|
| ak | 21,435 | 19.27% | 41.67% |
| dag | 5,307 | 7.04% | 23.75% |
| ee | 6,820 | 9.66% | 30.76% |
| en | 56,364 | 12.91% | 26.05% |
| ga | 9,621 | 8.14% | 15.83% |
| Combined | 99,547 | 13.66% | 29.19% |
Run locally
uvx --from huggingface-hub hf download Qlerqly/griot-nano-1 \
--local-dir griot-nano-1
cd griot-nano-1
# Transcribe an audio file
uv run --with-requirements requirements.txt python inference.py \
--model-dir . --audio recording.wav
# Live microphone transcription
uv run --with-requirements requirements.txt python inference.py \
--model-dir . --live
# Explicitly select BF16 or FP32 weights
uv run --with-requirements requirements.txt python inference.py \
--model-dir . --audio recording.wav --dtype bf16
--dtype auto is the default: the canonical FP32 weights are converted to BF16
on supported CUDA GPUs and remain FP32 elsewhere. Audio is converted to mono and
resampled to 16 kHz when necessary.
Recommended: KenLM decoding
For best contextual accuracy, we encourage using Griot Nano 1 together with Qlerqly/griot-nano-1-kenlm. KenLM adds n-gram beam-search guidance without another neural network, keeping memory and compute overhead low relative to neural language-model rescoring.
# From the directory containing griot-nano-1
uvx --from huggingface-hub hf download Qlerqly/griot-nano-1-kenlm \
--local-dir griot-nano-1-kenlm
cd griot-nano-1
uv run --with-requirements requirements-kenlm.txt python inference.py \
--model-dir . --audio recording.wav \
--kenlm ../griot-nano-1-kenlm/multilingual.arpa --beam-width 25
This runs the acoustic model and language model as one transcription pipeline.
The suggested starting values are beam width 25, LM weight 0.5, and word
insertion bonus 1.0; override the latter two with --lm-alpha and --lm-beta.
Greedy CTC remains available for the lowest latency. KenLM is most useful when word boundaries or acoustically similar character sequences are ambiguous.
Training data
The model was trained on multilingual Ghanaian speech covering Akan, Dagbani, Ewe, Ghanaian English, and Ga. See ATTRIBUTIONS.md for the complete dataset list, creators, source links, licenses, and usage notes.
License
The conservative release license is CC BY-NC-SA 4.0, not MIT. AfriSpeech
declares CC BY-NC-SA 4.0, configured GhanaNLP Community datasets declare CC
BY-NC 4.0, the University of Ghana WAXAL subsets and Ashesi FISD declare CC BY
4.0, and Common Voice is CC0. The audit is provided in license_audit.json.
This is a provenance-based engineering recommendation, not legal advice.
Limitations
- Akan is the hardest evaluated language and shows frequent
ɛ/e/a/iandɔ/o/uconfusions. - Word-boundary errors account for a material portion of Akan CER.
- Training data includes automatically generated or corrected transcripts.
- Performance outside the evaluated Ghanaian domains may differ.
Attributions
The complete training-data acknowledgements, source links, and license links are maintained in ATTRIBUTIONS.md.