SonoJot/sonojot-punct-v2

SonoJot Punctuation v2

This small, local model cleans up speech-recognizer output in two ways, in one forward pass:

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. v2 continues training from sonojot-punct-v1 and adds the span head.

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 0.17 s per 300 words when warm, on one desktop CPU with 2 threads (v1: about 0.16 s)
Training language	English

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 (a linear layer on the same hidden state).
  • span_labels.json: the span labels.

Input encoding

This is the same contract as v1:

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 decides the written form; the model does not:

• Zero to nine stay as words in prose.
• Digits from 10.
• Symbols for money and percent.
• Idioms and approximate amounts ("a hundred bucks", "twenty-four seven") stay as words.

SonoJot uses the span head for the numeric classes. AI model names, versions and AMP names come from a rules lexicon in the app, because the model tends to cut model names short. This combination measured best (below).

Results

Punctuation

Measured on 60 long real dictations (24,659 words), held out from training and scored against GPT-6.1 Sol references. Two annotators agree at sentence F1 0.940 and comma F1 0.907.

System	Sentence F1	Mid-clause false breaks / 1,000 words	Comma F1
Recognizer output	0.747	16.7	0.744
v1	0.888	0.8	0.876
v2	0.886	0.9	0.884

Sentence breaks are on par with v1. Commas are slightly better: +0.008, with a paired bootstrap 95% interval of +0.000 to +0.015.

Written form

The score is span F1 on held-out data: a span counts only if its boundaries, class and rendered text all match.

Test set	Rules only	Span head only	Span head + rules (SonoJot)
88 real dictations (GPT-6.1 Sol labels)	0.806	0.798	0.864
Earnings-21, real recognizer output	0.622	0.918	0.914
Earnings-22, spoken-form text	0.769	0.949	0.950
Synthetic dictation-style text	0.957	1.000	0.984

On the dictations, SonoJot's combination has 92% span precision. It changed one of the 50 dictations that needed no change.

Training data

Training ran for 3 epochs on an RTX 5070, starting from v1. The loss is masked punctuation cross-entropy plus masked span cross-entropy.

• Earnings-21 (Rev.com, revdotcom/speech-datasets, CC BY-SA 4.0). The calls were transcribed with Qwen3-ASR 1.7B through SonoJot's production pipeline.
  • Punctuation labels come from the professional references, normalized to the house style by GPT-6.1 Sol.
  • Span labels come from the dataset's entity tags and spoken-form normalizations, aligned to the recognizer's words. A span is accepted only when the recognizer's words match the written value.
• Earnings-22 (same source and licence). The spoken-form input is built from the references and their spoken normalizations. These rows train span labels only.
• Synthetic dictation-style text.
  • Built from 450 hand-written carrier sentences with typed slots: money, dates, times, versions, units, AI model names (including common recognizer mishearings) and AMP names.
  • 27% of the carriers are keep cases, where number words must stay as words.
  • The rows are checked against the held-out evaluation set for overlap (none share an 8-word sequence).
• Dictation. About 40,000 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 tends to keep long "and then … and then" chains as one sentence.
• Span coverage.
  • Few real dictation examples exist for times, ordinals and measures.
  • AI model names change quickly, which is why SonoJot renders them from a lexicon instead of trusting the span head.
• Domain. It is tuned for one speaker's dictation style, plus earnings-call speech.