Web Speech API: Unspoken Punctuation

concepts

1. Dictation Comparison

   Two recognisers running on the same mic. One with unspokenPunctuation: true, one without. See the same speech rendered two ways, side by side.

2. Code Mode

   The verbatim case the explainer carves out: coding by voice. With unspokenPunctuation: false the raw transcript surfaces "open paren", "comma" as literal words you can map to your own grammar — auto-inferred punctuation would break this.

3. Email Composer

   A realistic email editor where you dictate the body naturally. With unspokenPunctuation: true, commas and periods appear from your speech rhythm without you saying them aloud. Toggle the feature on/off mid-session to feel the difference. Live word and punctuation counts track how much the feature adds to your dictation.

4. Punctuation Analyzer

   Paste or type raw speech transcript and see exactly which commas, periods, and question marks the engine would infer. Five preset text samples — prose, email, list, code mode, interview — each highlight inferred punctuation in context. Counters track sentence boundaries, mid-sentence commas, and question inferences.

5. Settings Panel

   A full SpeechRecognition configuration builder with every property that affects punctuation inference: unspokenPunctuation, continuous, interimResults, lang, maxAlternatives. The generated JavaScript code block updates in real time; copy it directly into your project. Includes a live mic test using the built configuration.

6. Transcript Post-Processor

   Paste a raw speech transcript with no punctuation and watch a rule-based processor add commas, periods, and question marks — simulating what Chrome 150's unspokenPunctuation: true infers from prosody. A diff view highlights every added punctuation mark in green. Four presets (prose, email, list dictation, Q&A) show how the rules apply in different contexts. Stats track sentence boundaries, commas, and question marks inferred.

   Interactive NLP Diff view

why it shipped

Saying "comma" out loud feels unnatural. It's the single biggest complaint about dictation-driven text entry, and it's why most people abandon dictation for anything longer than a short message. Modern speech models can place punctuation by listening to prosody — when a sentence ends, the speaker's pitch drops; before a comma, there's usually a micro-pause. The on-device model already knows this; unspokenPunctuation just opts in to letting it act on the knowledge.

references

• ChromeStatus entry
• Web Speech API spec