Truecaser Models
Truecasing models for restoring proper capitalization in lowercase ASR output. Used by CrispASR via --truecase-model.
Available Models
| File | Type | Language | Size | F1 | License | Flag |
|---|---|---|---|---|---|---|
truecaser-lstm-de.bin |
BiLSTM char-level | German | 3.2 MB | 97.9% | Apache-2.0 | lstm or lstm-de |
truecaser-lstm-en.bin |
BiLSTM char-level | English | 3.2 MB | 93.0% | Apache-2.0 | lstm-en |
truecaser-lstm-es.bin |
BiLSTM char-level | Spanish | 3.2 MB | — | Apache-2.0 | lstm-es |
truecaser-lstm-ru.bin |
BiLSTM char-level | Russian | 4.1 MB | — | Apache-2.0 | lstm-ru |
truecaser-crf-de.bin |
CRF + context | German | 8.5 MB | ~95% | MIT | crf |
truecaser-de.bin |
Statistical freq | German | 1.7 MB | ~93% | MIT | auto |
BiLSTM Truecaser (recommended)
Converted from mayhewsw/pytorch-truecaser (Apache-2.0).
- Architecture: Embedding(202, 50) → BiLSTM(50→150, 2 layers) → Linear(300, 2)
- Labels: L (lowercase), U (uppercase) — per character
- Training: WMT monolingual text (de: 2.6M tokens, 97.86% F1; en: Wikipedia, 93.01% F1; es: WMT; ru: LORELEI)
- Original paper: Mayhew et al., "NER and POS When Nothing is Capitalized" (2019)
- Source: mayhewsw/pytorch-truecaser v1.0 —
wmt-truecaser-model-de.tar.gz
Example
Input: die schnelle braune katze springt über den faulen hund
Output: Die schnelle braune Katze springt über den faulen Hund
Correctly handles:
- Adjective vs noun: "braune" (lowercase) vs "Katze" (capitalize)
- Formal pronouns: "Ihnen" (capitalize)
- Compound words and proper nouns
CRF Truecaser
Trained on 245K sentences of WMT News Crawl German using python-crfsuite.
- Features: word identity, 3-char suffix, noun suffixes, previous/next word, article context
- Decode: Viterbi over linear-chain CRF (3 labels: lc, u1, uc)
- Training data: WMT News Crawl 2023 German (8.5 MB model, MIT license)
Statistical Truecaser
Simple word-frequency lookup trained on WMT News Crawl 2023 German.
- Entries: 71,142 unique words
- Size: 1.7 MB
- Approach: for each word, pick the casing variant (lowercase/capitalize/uppercase) seen most often
- Training data: WMT News Crawl 2023 German (278K sentences), MIT license
Usage with CrispASR
# BiLSTM (recommended)
crispasr --backend wav2vec2-de -m model.gguf --truecase-model lstm -f audio.wav
# CRF
crispasr --backend wav2vec2-de -m model.gguf --truecase-model crf -f audio.wav
# Statistical
crispasr --backend wav2vec2-de -m model.gguf --truecase-model auto -f audio.wav
# Combined with punctuation restoration
crispasr --backend moonshine -m model.gguf --punc-model punctuate-all --truecase-model lstm -f audio.wav
Conversion
# BiLSTM: download from mayhewsw, convert to binary
wget https://github.com/mayhewsw/pytorch-truecaser/releases/download/v1.0/wmt-truecaser-model-de.tar.gz
tar xzf wmt-truecaser-model-de.tar.gz
python models/convert-lstm-truecaser-to-bin.py --input wmt-truecaser-de/ --output truecaser-lstm-de.bin
# CRF: train from Wikipedia
python models/train-truecaser-crf.py --output truecaser-crf-de.bin
Provenance — mixed authorship, and outside the GPAI definition
Added 2026-08-02 during an account-wide provenance review.
This repository holds weights of two different origins, which is why the
per-file table above matters more here than in a typical conversion repo:
the BiLSTM truecasers were converted from mayhewsw/pytorch-truecaser (Apache-2.0), while the CRF and statistical truecasers were trained here on WMT News Crawl German. Conversion and training are different acts with different attribution consequences, and the table above is what records which file is which.
Most cstr/* repositories are conversions where the upstream research team
remains the provider of the model. That description is only half true here, and
a blanket "format conversion only" statement — of the kind added across the
conversion repos in this review — would have been inaccurate for this one. It is
deliberately not made.
EU AI Act Art. 53 does not apply. Art. 53 binds providers of general-purpose AI models, defined by Art. 3(63) as models displaying "significant generality" and capable of "competently performing a wide range of distinct tasks". A truecaser restoring capitalisation does not meet that definition, so the Art. 53(1)(c) copyright-policy and 53(1)(d) training-content duties are not engaged for either set of weights. The training-data and licence documentation above is published because attribution is required and because it is useful — not because Art. 53 compels it.