waxal-kenlm-models

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 CTC acoustic models. Each model was trained as part of a Zindi ASR competition workflow on the WaxalNLP benchmark.

These models are trained on normalized text — lowercased, with training targets restricted to a fixed alphabetic character set. This is the companion repo to keystats/waxal-kenlm-models-best, which was built the same way but on raw (cased, punctuated) text. The two repos are not interchangeable — see Pairing below.

Files

Each language folder contains:

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

Normalization

Text was lowercased and normalized (NFKC) before being added to the training corpus, matching the tokenizer/vocabulary used by the paired -main acoustic models (character-level, lowercase, | as word delimiter, no punctuation). This keeps the LM's vocabulary aligned with what those specific acoustic models can output — they were trained to predict lowercase, unpunctuated text by design, not retrained for casing the way the -main-best siblings were.

Pairing — important

Use the matching KLM variant for whichever ASR checkpoint you're decoding with:

Acoustic model	Text convention	Matching KLM repo
keystats/w2v-bert-2.0-{lingala,shona,luganda}-main	lowercase, no punctuation	this repo (waxal-kenlm-models)
keystats/w2v-bert-2.0-{lingala,shona,luganda}-main-best	raw, case + punctuation preserved	keystats/waxal-kenlm-models-best

Mixing a lowercase-only ASR model with the raw-text KLM (or vice versa) causes a vocabulary mismatch during shallow-fusion decoding and will hurt accuracy rather than help it.

Training data (per language)

Same underlying source pools as the -best companion repo — only the text normalization differs. Corpora were built exclusively from train-split / non-benchmark text; WaxalNLP validation and test splits were 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 (lowercased) 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 CTC model to do beam-search decoding with shallow fusion. Output will be lowercase and unpunctuated, consistent with the acoustic model's training targets — apply your own capitalization/punctuation post-processing if you need readable output, or switch to the -best model + KLM pair for cased/punctuated output directly from decoding.

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.
• Lowercasing and dropping punctuation discards information (sentence boundaries, proper-noun casing) that the raw-text -best variant retains.
• 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, keystats/w2v-bert-2.0-shona-main, keystats/w2v-bert-2.0-luganda-main
• Raw-text companion repo: keystats/waxal-kenlm-models-best
• Benchmark: google/WaxalNLP