concept · Web API · Speech

Quality Level Explorer

Select a recognition quality tier, then speak. The quality property tells the browser which on-device model to prefer — a lightweight command model for short utterances, a full dictation model for long-form speech, or a conversation model for noisy multi-speaker environments.

Your browser does not support the Web Speech API (SpeechRecognition). Try Chrome 150+ on desktop.

Recognition language on-device only

Selected language: English (United States), BCP 47 tag en-US. Availability has not been checked yet.

Fast

'command'

Short phrases, single speaker, limited vocabulary. Lowest resource use — ideal for voice controls and quick queries.

Balanced

'dictation'

Continuous speech, moderate noise, single speaker. Good for note-taking, email, and long-form text entry.

High quality

'conversation'

Multi-speaker, complex vocabulary, high noise tolerance. Meeting transcription and live captions.

quality: 'command' · processLocally: true

Transcript command

Select a quality level and click "Start recognizing"…

The quality property is a hint, not a guarantee. If the device lacks the hardware for the requested tier, the browser will fall back to the best available model. The hint allows the browser to make a more informed selection rather than always defaulting to the most lightweight option, which previously caused accuracy regressions for apps that needed higher fidelity.

the API

const recognition = new SpeechRecognition();

// Request on-device processing
recognition.processLocally = true;
recognition.continuous     = true;
recognition.interimResults = true;
recognition.lang           = 'en-US';

// New in Chrome 150: tell the browser which quality tier you need
recognition.quality = 'dictation'; // 'command' | 'dictation' | 'conversation'

recognition.onresult = (event) => {
  const transcript = [...event.results]
    .map(r => r[0].transcript)
    .join('');
  updateUI(transcript);
};

implementation reference

Need the exact API surface, compatibility boundaries, errors, lifecycle, and source links? Read the matching gendn reference ↗

the API

see also

implementation reference