waxal-kenlm-models-best

5-gram KenLM language models for Lingala (ln), Shona (sn), and Luganda (lg), built for shallow-fusion decoding (via pyctcdecode) alongside the corresponding keystats w2v-bert-2.0-*-main-best CTC acoustic models. Each model was trained as part of a Zindi ASR competition workflow on the WaxalNLP benchmark.

These models are trained on non-normalized, raw text — case and punctuation are preserved, not lowercased or stripped. This is a deliberate choice, explained below.

Files

Each language folder contains:

File	Description
*_5gram-best.klm	Binary KenLM model (trie format), ready for use with pyctcdecode / kenlm.Model.
*_5gram_correct-best.arpa	Human-readable ARPA file the binary was built from, with a </s> unigram entry added (see below).

Why non-normalized text?

An earlier version of this pipeline lowercased text and stripped it down to a fixed [a-zà-ÿɔɛ' ]-style character class before building the KenLM corpus. That was a bug: the paired acoustic models (keystats/w2v-bert-2.0-{lingala,shona,luganda}-main-best) were retrained to predict case and punctuation, and Zindi's own scoring is raw WER/CER with no normalization on either side. A lowercased, punctuation-stripped LM actively fights a case-and-punctuation-aware acoustic model during shallow fusion, pulling correct predictions back toward wrong ones.

These models fix that: the allowed character set for the training corpus is derived directly from each acoustic model's tokenizer vocabulary (letters of any case, apostrophes, and punctuation the tokenizer can actually emit), rather than a hand-picked regex. Text is NFKC-normalized and any character outside that derived set is dropped, but case and punctuation are kept intact throughout. This keeps the LM's vocabulary aligned with what the CTC model can actually output at decode time.

Training data (per language)

Corpora were built exclusively from train-split / non-benchmark text — WaxalNLP validation and test splits were explicitly excluded from every corpus to avoid leaking evaluation data into the decoder.

Lingala — pooled from:

• KasuleTrevor/lingala_10hr
• Svngoku/lingala-asr-dataset
• shunyalabs/lingala-speech-dataset
• KasuleTrevor/Lingala_100hrs (all splits — these are the source's own splits, not the WAXAL benchmark)
• WaxalNLP Lingala train split only
• Regineforte/asr_lingala

Shona — pooled from:

• WaxalNLP Shona train split only
• badrex/shona-speech
• Beijuka/DigitalUmuganda_AfriVoice_shona
• shunyalabs/shona-speech-dataset
• realtime-speech/shona2

Luganda — pooled from:

• WaxalNLP Luganda train split only
• keystats/luganda_asr_dataset
• FarmerlineML/luganda_dataset_2.0
• Bateesa/luganda-tts-toby

Exact-duplicate lines were removed after normalization (order-preserving) so no single overlapping source could over-weight the LM.

Training procedure

• Toolkit: KenLM (lmplz + build_binary)
• Order: 5-gram
• Command: lmplz -o 5 -S 40% --discount_fallback over the normalized text corpus
• Post-processing: lmplz does not emit a </s> unigram entry by default, which pyctcdecode requires — the unigram count and an </s> line (duplicated from <s>) were added to the ARPA file before compiling to binary (hence the _correct filename suffix).
• Binary format: compiled with build_binary trie for memory-efficient loading.

Intended use

Load a .klm file into pyctcdecode.build_ctcdecoder(..., kenlm_model_path=...) alongside the matching keystats/w2v-bert-2.0-{lingala,shona,luganda}-main-best CTC model to do beam-search decoding with shallow fusion. Because both the acoustic model and this LM predict case and punctuation, do not lowercase or strip punctuation from hypotheses before scoring — doing so no longer matches how these models were built or how Zindi scores submissions.

Limitations

• Corpora mix ASR-transcript text (informal, sometimes noisy) with other web/community-sourced text; quality and register vary by source.
• Not deduplicated against near-duplicates (only exact-string duplicates were removed), so some sources may still be over- or under-represented.
• No manual QA of transcript accuracy in the upstream source datasets — errors in those sources propagate into the LM.
• License/usage terms of upstream source datasets vary and are not unified here; check each source's own license before downstream/commercial use.

Related

• Acoustic models: keystats/w2v-bert-2.0-lingala-main-best, keystats/w2v-bert-2.0-shona-main-best, keystats/w2v-bert-2.0-luganda-main-best
• Benchmark: google/WaxalNLP