cstr/fullstop-punc-multilang-GGUF

Fullstop Punctuation Multilingual (GGUF)

GGUF conversion of oliverguhr/fullstop-punctuation-multilang-large for use with CrispASR.

Adds punctuation to unpunctuated ASR output. Multilingual (English, German, French, Italian) with ASCII punctuation output.

Model Details

• Architecture: XLM-RoBERTa-large — 24L, d=1024, 16 heads, d_ffn=4096, GELU
• Parameters: ~560M
• Classifier: Linear(1024, 6) — 6 punctuation classes
• Labels: none, . (period), , (comma), ? (question), - (dash), : (colon)
• Vocabulary: SentencePiece (250,002 tokens)
• Max sequence: 512 tokens (auto-chunked)
• Languages: English, German, French, Italian
• License: MIT

Usage with CrispASR

crispasr --backend wav2vec2 -m wav2vec2.gguf --punc-model fullstop-punc-q4_k.gguf -f audio.wav

Available Files

File	Quant	Size	Description
fullstop-punc-f32emb.gguf	F16+F32emb	1.6 GB	Full precision (F32 embeddings)
fullstop-punc-q8_0.gguf	Q8_0	572 MB	High quality
fullstop-punc-q4_k.gguf	Q4_K	254 MB	Recommended

Example

Input: and so my fellow americans ask not what your country can do for you ask what you can do for your country

Output: And so my fellow americans ask not what your country can do for you, ask what you can do for your country.

Conversion

python models/convert-fullstop-punc-to-gguf.py \
  --input oliverguhr/fullstop-punctuation-multilang-large \
  --output fullstop-punc.gguf

Original Model

• Source: oliverguhr/fullstop-punctuation-multilang-large
• Paper: fullstop-deep-punctuation-prediction
• Training: Europarl Dataset (political speeches)

Provenance and EU AI Act Art. 53 note

• Upstream model: oliverguhr/fullstop-punctuation-multilang-large — published by oliverguhr.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.