SonoJot/sonojot-punct-v1

SonoJot Punctuation v1

This small, local model repairs punctuation in speech-recognizer output. SonoJot uses it after Qwen3-ASR to remove the false sentence breaks that recognizers put at hesitation pauses ("The problem is that. I need it"). It also places commas, question marks and the occasional semicolon.

It is a token classifier. For each word it predicts the mark that follows: none, ., ,, ? or ;. It reads the recognizer's own punctuation as evidence, so it keeps real sentence ends and removes pause artifacts. It never adds, removes or changes words.

Base model	Horizon-Labs/punctuation-restoration-small (mmBERT-small, 140M parameters)
Runtime file	model.int8.onnx, 268 MB: int8 embeddings, fp32 compute (full int8 costs about 2 F1 points)
Speed	About 0.5 ms per word on one desktop CPU with 2 threads (about 0.17 s per 300 words when warm)
Training language	English

Files

• model.int8.onnx: inputs input_ids and attention_mask (int64, [1, T]); output logits ([1, T, 5]).
• tokenizer.json: Hugging Face tokenizers format.
• punct-config.json: labels, window and hop sizes, special-token ids, and the exact input encoding.
• manifest.json: byte sizes and SHA-256 of the three runtime files. SonoJot verifies these before loading.
• pytorch/: the fine-tuned safetensors weights and config, for further fine-tuning.

Input encoding

1. Split the text into words ([^\W_]+(?:['’][^\W_]+)*) and lowercase each one, except I, I'm, I'll, I've and I'd.
2. For each word, append the tokens of " " + word. If the recognizer wrote a mark after the word, also append that mark's tokens (. for ./!, ?, or , for ,;: and dashes; glued marks such as 1.2 and Next.js don't count). The last word always carries a mark.
3. Wrap the sequence in CLS … SEP. Read each word's label at its last word token.
4. Long texts use windows of 120 words with a hop of 60. For each word, keep the prediction from the window where it is farthest from an edge.

punct-config.json repeats this contract.

Results

Test set. 60 long real dictations (24,659 words), produced by Qwen3-ASR followed by SonoJot's local cleanup rules. None were used in training. They were scored against an independent GPT-6.1 Sol reference made under the house style below. Two annotators agree with each other at sentence F1 0.940 and comma F1 0.907.

System	Sentence-boundary F1	Mid-clause false breaks / 1,000 words	Comma F1
Recognizer output	0.747	16.7	0.744
Qwen3-4B-Instruct rewrite (guarded)	0.859	4.0	0.764
Gemma 4 E4B rewrite (guarded)	0.881	2.8	0.785
This model	0.887	0.8	0.876

Held-out Earnings-21 calls (real recognizer output): sentence F1 0.895. Synthetic held-out documents with dictation-rate recognizer errors: F1 0.928, comma F1 0.891.

House style

The labels follow SonoJot's house style:

• No dashes or ellipses.
• Commas separate restarts and asides ("put it on your, your sheet").
• Semicolons appear only between two complete clauses spoken as one thought, with no conjunction between them.
• Every sentence end is followed by a capital letter. The app applies capitalization and dash removal; the model predicts the marks.

Training data

Training ran for 3 epochs on an RTX 5070.

• Earnings-21 (Rev.com, revdotcom/speech-datasets, transcripts CC BY-SA 4.0). 41 of the 44 calls were transcribed with Qwen3-ASR 1.7B through SonoJot's production pipeline. The professional reference punctuation was normalized to the house style by GPT-6.1 Sol, then projected onto the recognizer's words through alignment.
• Synthetic errors. The same references, corrupted with false periods, dropped commas and missed sentence ends, at rates measured on real dictation.
• Dictation. About 39,000 words of one consenting user's own dictation, labeled by GPT-6.1 Sol. This text is not distributed.

License and attribution

The weights are released under CC BY-SA 4.0, because the training labels derive from the Earnings-21 transcripts (CC BY-SA 4.0, © Rev.com). The base model is Apache-2.0 (Horizon-Labs, built on jhu-clsp/mmBERT-small, MIT). If you redistribute this model or derivatives, keep this attribution and the share-alike terms.

Limitations

• Language. English only. Other languages pass through with weaker results.
• Casing. The model predicts punctuation, not casing; SonoJot capitalizes with rules.
• Run-on sentences. It tends to keep long "and then … and then" chains as one sentence (about 5 missed breaks per 1,000 words against the reference).
• Domain. It is tuned for one speaker's dictation style plus earnings-call speech.