demo · v150
Quality Latency Tradeoff
The three quality tiers — 'command', 'dictation', 'conversation' — trade recognition latency against accuracy and model capability. Pick a tier to see its expected latency budget, word-error rate range, resource cost, and a simulated transcript that shows the kinds of errors and improvements each tier produces. Then use the decision wizard to find the right tier for your use case.
Select a quality tier
command
Command
Latency: 50–150 ms
Accuracy: ~85%
RAM: ~30 MB
dictation
Dictation
Latency: 200–500 ms
Accuracy: ~95%
RAM: ~120 MB
conversation
Conversation
Latency: 500–1200 ms
Accuracy: ~97%
RAM: ~350 MB
P50 Latency
80
ms
Word Error Rate
15
%
Max Utterance
10
words
Speaker Sep.
No
diarization
Simulated transcript — same audio, different quality tier
Recognition error
Punctuation / correction added
Speaker label
Speed vs accuracy across tiers
Decision wizard — find the right tier
1. What type of speech input does your app handle?
2. How sensitive is the user to recognition latency?
3. What is the background noise environment?
4. How important is punctuation accuracy?
Recommendation
// Chrome 150: set quality on SpeechRecognition
const sr = new SpeechRecognition();
sr.quality = 'command'; // 50-150ms, ~85% accuracy — smart home, shortcuts
sr.quality = 'dictation'; // 200-500ms, ~95% accuracy — note-taking, emails
sr.quality = 'conversation'; // 500-1200ms, ~97%, multi-speaker — meetings
// Graceful fallback: if quality not supported, the property is silently ignored
// and the browser picks its default model
if ('quality' in SpeechRecognition.prototype) {
sr.quality = 'dictation';
}
sr.start();
see also
- Command vs. Dictation — two sessions side by side
- Use-Case Sampler — preset scenarios with correct tier
- Device Capability Probe — full readiness check
- Quality Config Panel — build a full SpeechRecognition config
- ChromeStatus: On-Device Recognition Quality
implementation reference
Need the exact API surface, compatibility boundaries, errors, lifecycle, and source links? Read the matching gendn reference ↗