SonoJot/sonojot-punct-v3

SonoJot Punctuation v3

v3 has the same architecture, files and input contract as sonojot-punct-v2. It is retrained with about twice as much real dictation, and dictation inputs are now built exactly as SonoJot builds them at run time.

In one forward pass, it cleans up speech-recognizer output in two ways:

1. Punctuation. 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 the false breaks that recognizers put at hesitation pauses ("The problem is that. I need it").
2. Written-form spans. It tags spans that a person would type differently from how they are spoken: "four thousand dollars" (MONEY), "ten percent" (PERCENT), "twenty twenty five" (YEAR), "ten K" (MAGNITUDE). The model only tags. A deterministic renderer in SonoJot writes $4,000, 10%, 2025 and 10K. It also checks that the written form reads back as the spoken words, and otherwise keeps them.

The model never adds, removes or changes words by itself.

Base model	sonojot-punct-v1 ← Horizon-Labs/punctuation-restoration-small (mmBERT-small, 140M parameters)
Runtime file	model.int8.onnx, 268 MB: int8 embeddings, fp32 compute
Speed	About 0.14 s per 300 words when warm, on one desktop CPU with 2 threads (release build, Ryzen 9 9900X; v2 measured the same way: 0.14 s)
Training language	English

What changed from v2

• More real dictation. 310 more dictations (about 56,000 words) from one consenting user, labeled by GPT-6.1 Sol under the same house style. That roughly doubles the in-domain data.
• Correct inputs. Each dictation input is rebuilt from the raw recognizer text with SonoJot's rules alone. Since v2 shipped, the saved text already carries v2's punctuation, so it can't serve as input.
• Training. v3 retrains from v1 with v2's recipe on v2's data plus the new rows.
  • The recipe: 3 epochs, AdamW at 4e-5, effective batch 16, masked punctuation plus span loss, and new dictation windows weighted 5x.
  • Training ran three times with different seeds. The released run is the one with the best development score, not the best test score.

Files

• model.int8.onnx:
  • Inputs: input_ids and attention_mask (int64, [batch, T]).
  • Outputs: logits ([batch, T, 5], punctuation) and span_logits ([batch, T, 27], BIO span labels).
• tokenizer.json: Hugging Face tokenizers format.
• punct-config.json: punctuation labels, span 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 weights for further fine-tuning:
  • model.safetensors and config.json: the encoder and the punctuation head.
  • span_head.pt: the span classifier.
  • span_labels.json: the span labels.

Input encoding

This is the same contract as v1 and v2:

1. Split the text into words ([^\W_]+(?:['’][^\W_]+)*). Lowercase each word, except I, I'm, I'll, I've and I'd.
2. For each word, append the tokens of " " + word.
3. If the recognizer wrote a mark after the word, also append that mark's tokens: . for . and !, ? for ?, or , for ,, ;, : and dashes. The last word always carries a mark.
4. Wrap the sequence in CLS … SEP.
5. Read both heads at each word's last word token. Take the argmax of each head independently.
6. 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.

Span classes

CARDINAL, DECIMAL, MONEY, PERCENT, MAGNITUDE (10K, 4K, 10x), YEAR, DATE, TIME, ORDINAL, MEASURE (number + technical unit), VERSION, MODEL (AI model names), AMP (R&D-style names), in BIO form (27 labels).

The house style is unchanged from v2. SonoJot uses the span head for the numeric classes. AI model names, versions and AMP names come from a rules lexicon in the app.

Results

Punctuation

The test set is 88 real dictations (29,177 words), held out from training. Inputs are rules-only recognizer text, as SonoJot produces at run time.

• Reference: GPT-6.1 Sol labeled the test set. An Opus 5.5 annotator then labeled it blind, and separate judges settled every disagreement under the house style.
• Rows: the v2 and v3 rows score the published or released models. "Mean of 3 seeds" trains each recipe three times.

System	Sentence F1	False breaks / 1,000 words	Missed breaks / 1,000 words	Comma F1
v2 (published)	0.882	5.9	5.4	0.881
v2 recipe, mean of 3 seeds	0.883	6.3	5.0	0.878
v3 recipe, mean of 3 seeds	0.894	5.9	4.5	0.884
v3 (released seed)	0.897	6.0	4.1	0.884

• Ranking: every v3 training run beat every v2 training run on sentence F1 and comma F1.
• Capitalization errors fell from 9.9 to 9.2 per 1,000 words (seed means).
• Ceiling: two careful annotators agree at about 0.94 sentence F1 and 0.92 comma F1. That is roughly the most any model can score against a single reference.

Written form

The score is span F1 for SonoJot's combination of the span head and rules, on held-out data. A span counts only if its boundaries, class and rendered text all match.

Test set	v2	v3
88 real dictations	0.870	0.848
Earnings-21, real recognizer output	0.912	0.910
Earnings-22, spoken-form text	0.947	0.945
Synthetic dictation-style text	0.984	0.983

• Overall: written form is about the same as v2. Another v3 training run scored 0.872 on the dictations, so the difference is within seed variance on 113 gold spans.
• Challenge set: on the 763-case written-form challenge, SonoJot's combination gets 702 exact with v3 against 694 with v2. The span head alone gets 663 against 639.

Training data

The data is v2's (see the v2 card) plus the new dictations:

• Earnings-21 and Earnings-22 (Rev.com, revdotcom/speech-datasets, CC BY-SA 4.0). The calls were transcribed with Qwen3-ASR 1.7B through SonoJot's production pipeline. Labels come from the professional references, normalized to the house style by GPT-6.1 Sol.
• Synthetic dictation-style text for written-form spans, unchanged from v2.
• Dictation. About 95,000 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 and Earnings-22 transcripts (CC BY-SA 4.0, © Rev.com). The base model chain 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.
• Casing. The model predicts marks and spans, not casing; SonoJot capitalizes with rules.
• Run-on sentences. It still tends to keep long "and then … and then" chains as one sentence.
• Span coverage. Few real examples exist for times, ordinals and measures. AI model names come from the app's lexicon.
• Domain. It is tuned for one speaker's dictation style, plus earnings-call speech.