Web Speech API: On-Device Recognition Quality

concepts

1. Quality Level Explorer

   An interactive demo where you pick a quality tier and then speak. See how the transcript differs across 'command' (fast, low-resource), 'dictation' (accurate continuous speech), and 'conversation' (multi-speaker, high-noise) settings.

2. Command vs. Dictation

   Run two SpeechRecognition sessions simultaneously — one with quality: 'command' for snappy short responses, one with quality: 'dictation' for long-form accuracy — and compare the transcripts side by side.

3. Use-Case Sampler

   Three preset real-world scenarios — smart home control, note taking, meeting transcription — each with its correct quality tier pre-configured. Select a scenario to see the full SpeechRecognition configuration and try live recognition with that quality level applied.

4. Quality Config Panel

   Build a SpeechRecognition configuration from scratch. Choose a quality tier, toggle continuous, interimResults, unspokenPunctuation, and other options — the generated JavaScript code updates live. Copy the snippet directly into your project, or click to test live recognition with the built config.

5. Device Capability Probe

   A six-point capability check: SpeechRecognition API presence, the quality property (Chrome 150), on-device availability (Chrome 149), unspokenPunctuation, microphone permission state, and secure-context status. Run the probe to get a full readiness report for on-device speech recognition on this device.

6. Quality Latency Tradeoff

   Visualises the speed-vs-accuracy tradeoff across all three quality tiers: 'command' (50–150 ms, ~85% accuracy), 'dictation' (200–500 ms, ~95%), and 'conversation' (500–1200 ms, ~97%, multi-speaker). Simulated transcripts for the same audio show how each tier handles errors, punctuation, and speaker labels. A four-question decision wizard recommends the right tier for your use case.

   Interactive Latency Decision tree

why it shipped

The on-device Web Speech API (Chrome 149) exposes speech recognition that runs locally without a cloud round-trip. But on-device models come in tiers: a lightweight command-recognition model is fast and fits on most devices, while a high-quality dictation model is more accurate but requires more RAM and compute. Before Chrome 150 the browser always picked the model silently. With the quality hint, a smart-home voice-control app can opt for the lighter command model to minimise latency, while a note-taking app can opt for the dictation model to improve transcript fidelity. Google Meet uses this property to guarantee its required accuracy tier before relying on on-device recognition.

quality levels

• 'command' — short phrases, single speaker, limited vocabulary (e.g., "turn off the lights", "next slide"); lowest resource use, fastest response
• 'dictation' — continuous speech, moderate background noise, single primary speaker (e.g., long-form text entry, note-taking, email dictation); medium resource use
• 'conversation' — multi-speaker, complex vocabulary, high noise tolerance (e.g., meeting transcription, live captions); highest resource use, requires capable hardware

references

• ChromeStatus entry
• MDN — SpeechRecognition
• On-device speech recognition explainer
• Web Speech API specification (WICG)