Griot Nano 1 Multilingual KenLM
This is the multilingual 3-gram KenLM language model for Qlerqly/griot-nano-1. It covers Akan, Dagbani, Ewe, Ghanaian English, and Ga using the same normalized transcript domain as the Griot Nano 1 acoustic model.
Recommended decoding configuration
For best contextual accuracy, we recommend KenLM-assisted CTC beam search with Griot Nano 1. It improves the decoder's preference for plausible character and word sequences, particularly around ambiguous acoustics and word boundaries.
KenLM is a compact n-gram model rather than a second neural network. This keeps memory and compute overhead low relative to neural language-model rescoring. Greedy CTC remains the lowest-latency option.
Suggested starting settings:
- Beam width:
25 - LM weight (
alpha):0.5 - Word insertion bonus (
beta):1.0
These are starting points rather than universal optima; tune them on a held-out set for your target language and domain.
Results Vs Greedy and Beam search decode
| Language | Samples | Greedy CER | KenLM CER | Greedy WER | KenLM WER | Relative WER reduction |
|---|---|---|---|---|---|---|
Akan (ak) |
21,538 | 19.30% | 19.60% | 41.67% | 36.43% | 12.58% |
Dagbani (dag) |
5,347 | 7.05% | 5.64% | 23.17% | 14.56% | 37.16% |
Ewe (ee) |
6,939 | 10.01% | 9.22% | 30.97% | 24.60% | 20.57% |
English (en) |
56,592 | 12.76% | 12.45% | 25.38% | 20.93% | 17.56% |
Ga (ga) |
9,707 | 8.65% | 8.40% | 16.02% | 13.12% | 18.12% |
| Combined | 100,123 | 13.58% | 13.33% | 28.73% | 23.91% | 16.76% |
Download and use
uvx --from huggingface-hub hf download Qlerqly/griot-nano-1-kenlm \
--local-dir griot-nano-1-kenlm
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
The acoustic-model repositories provide requirements-kenlm.txt and the
KenLM-enabled inference script.
Model details
- Format: ARPA
- Order: 3-gram
- Size: 214.2 MiB
- SHA-256:
c017eaaa575a0bb0dfc30ab74953a95f9e11939ca7d0a034eed3c446ded328f6 - Quantized: no
| N-gram order | Entries |
|---|---|
| 1 | 524,748 |
| 2 | 2,215,703 |
| 3 | 4,933,206 |
Training data and license
This language model was derived from normalized transcripts associated with the Griot Nano 1 training sources. See ATTRIBUTIONS.md for the complete source, creator, and license record.