Brave on-device speech recognition test page

This plan tests the three Web Speech API calls for on-device speech recognition:

• available() asks if on-device speech works for a language.
• install() lets a website use the speech model. Brave asks the user before the first download.
• start() listens and returns text, from an audio track or the microphone.

Only English is supported. The plan does not test every option. It tests the real model download, the real model, a real microphone and real websites.

Before you start

• Use Brave 1.100.44 or later.

• Test page: open https://yrliou.github.io/speech-test-page/. Each button shows the code it runs. The log shows the result.

• Profile: use a brand-new profile. On macOS:

  "/Applications/Brave Browser Nightly.app/Contents/MacOS/Brave Browser Nightly" \
    --user-data-dir=/tmp/brave-speech-qa
  

  Windows and Linux take the same --user-data-dir option.

• Feature flag: brave://flags/#brave-on-device-speech-recognition. It is off by default. Section 2 turns it on.

• Model: 636 MB. Brave downloads it only after the user allows it.

Do the sections in order, on one profile. Section 6.2 needs a standalone Brave Origin build and a Brave Origin upgrade. Use a new profile for each, with a different --user-data-dir.

1. Feature flag off

Delete /tmp/brave-speech-qa, then launch with the command above.

1. Open the test page. Status says unavailable.
2. In 2. start(), keep Sample clip, tick processLocally and press Start. The log shows error language-not-supported.
3. Untick processLocally and press Start. The log shows error network, and no interim or final lines.

2. Feature flag on

1. Open brave://flags/#brave-on-device-speech-recognition. Brave On-Device Speech Recognition shows Default (Disabled). Set it to Enabled and press Relaunch.
2. brave://components has no "Brave On-Device Speech Models" entry.
3. Reload the test page. Status says downloadable.
4. Tick processLocally and press Start. The log shows error language-not-supported.

3. install()

This section uses only Allow. The test plan in brave-core#40814 covers the prompt and the settings switch.

1. In 1. available() and install(), press install(). A prompt opens. Press Allow. brave://components lists the model. The download can take a few minutes.

2. When the download is done, the log shows true. Status says available.

3. Change one option at a time in 1. available() and install(). Reset it after each row.

   Option	available()	install()
   quality dictation	'available'	true
   quality conversation	'unavailable'	false
   langs en-GB	'available'	true
   langs fr-FR	'unavailable'	false

   Brave's model serves the command and dictation qualities, and English only.

4. start()

Use Sample clip, unless the step says Microphone. Reset the options to the defaults after each row.

The sample clip plays through an audio track at the computer's audio rate. The first log line shows it, usually 44100 or 48000 Hz. The model takes 16000 Hz, so this also tests the audio conversion.

processLocally

processLocally	Expected in the log
ticked	interim lines, then final "This is a sentence in a single segment"
unticked	The same

Options

Tick processLocally for each row.

Option	Expected in the log
continuous unticked	Ends by itself a short time after the speech ends
interimResults unticked	Only the final line
quality dictation	final with the sentence
quality conversation	error language-not-supported
lang fr-FR	error language-not-supported

Microphone

Pick Microphone and tick processLocally.

1. Press Start and allow the microphone. Say a sentence and press Stop. The final line shows your words.
2. Untick continuous and do it again. It stops by itself a short time after you stop talking.

Real websites

These websites do not set processLocally. They show your words:

• https://www.google.com/intl/en/chrome/demos/speech.html: choose English, United States. Click the microphone, speak, click it again.
• https://speechnotes.co/dictate/: click the microphone, speak, click it again.

5. After a relaunch and on a new origin

1. Quit Brave and launch it again with the same command. Reload the test page. Status says available. Tick processLocally and press Start. The log shows the sentence.
2. Open https://example.com, which is a different origin, and the DevTools console. Typing in the console counts as a click. Run:
   • await SpeechRecognition.available({ langs: ['en-US'], processLocally: true }) returns "downloadable". Each origin must call install() once.
   • await SpeechRecognition.install({ langs: ['en-US'], processLocally: true }) returns true right away, with no prompt.
   • await SpeechRecognition.available({ langs: ['en-US'], processLocally: true }) returns "available".

6. Other windows and builds

6.1 Private, Guest and Tor windows

Use the same profile. The model is on disk, but these windows must not use it. Open a New Private Window, a Guest window and a New Private Window with Tor. In each, open the test page and check:

Check	Expected
Status	unavailable
install()	false, with no prompt
Start, processLocally ticked	error language-not-supported
Start, processLocally unticked	error network, and no text

6.2 Brave Origin builds

Brave Origin comes in two types:

• Standalone: the separate Brave Origin app. It does not include on-device AI, so it must show no part of this feature.
• Upgrade: Brave with a Brave Origin purchase. Only this type has the brave://settings/origin page, with an On-Device AI switch.

Standalone

Use a brand-new profile. This build has no feature flag to turn on.

1. Make sure that brave://settings/privacy has no On-device speech recognition switch.
2. Open the test page:
   • Status says unavailable.
   • Start with processLocally ticked gives error language-not-supported.
   • Start with processLocally unticked gives error network, and no text.

Upgrade

The On-Device AI switch starts off. Use a brand-new profile, and turn on the feature flag as in section 2. Reload the test page after each change.

1. With the switch off:
   • Status says unavailable.
   • Start with processLocally ticked gives error language-not-supported.
   • Start with processLocally unticked gives error network, and no text.
2. Turn On-Device AI on. Status says downloadable. Press install(), then Allow. When the download is done, the log shows true. Start with processLocally ticked shows the sentence.
3. Turn On-Device AI off. The model disappears from brave://components. Status says unavailable. Start with processLocally unticked gives error network, and no text.

Credits

wpt_speech.wav is from web-platform-tests. See wpt_speech.wav.LICENSE.md.