Repository navigation
Replies: 1 comment
|
There is one qualification to the stuck-button explanation. In your checkout, clicking a speaking button calls Newer code at On a Listen click, first check |
|
There is one qualification to the stuck-button explanation. In your checkout, clicking a speaking button calls Newer code at On a Listen click, first check |
Uh oh!
There was an error while loading. Please reload this page.
WebUI version:
exp-v0.52.379(master,ff26335b)I can't get the built-in ElevenLabs engine to play. Server side looks correct to me, so I
suspect I'm missing a setup step — or there's a client bug. Details below.
Symptom
Settings → TTS Engine = ElevenLabs. Clicking the per-message Listen (speaker)
button on an assistant message produces no audio and no error toast, and repeated
clicks do nothing.
The same button works in this browser when a locally-registered custom engine
(
window.registerHermesTtsEngine) is selected — so the button, the handler and theplayback path are known-good on this machine.
Configuration
ELEVENLABS_API_KEYis in~/.hermes/.envand is present in the WebUI server processenvironment.
~/.hermes/webui/settings.jsonhastts_engine: "elevenlabs"andtts_enabled: true, and the server reportstts_engineinpersisted_speech_keys, so_mirrorSpeechSettingsFromServer(static/boot.js:3230) should be writinglocalStorage["hermes-tts-engine"] = "elevenlabs".Server side — verified working
I reproduced exactly the upstream call
_handle_ttsmakes (api/routes.py:20621→api/routes.py:20658):So the key, the
voice_id, the model and the upstream endpoint are all fine, and theroute resolves
config.yamlcorrectly (voice_id, andmodelbeforemodel_id).Client path traced
speakMessage(static/ui.js:9508) readslocalStorage['hermes-tts-engine'], dispatchesto
_playElevenLabsTts(static/ui.js:9572) which POSTs/api/tts, then hands theArrayBuffer to
_playAudioBuf(static/ui.js:9638) →decodeAudioData→AudioBufferSourceNode.start(0).Every failure path there appears to call
showToast(...), so a fully silent click isunexpected. I could not capture browser console output — the WebUI password in
~/.hermes/.envis a PBKDF2 salt (api/auth.py), so I can't log in programmatically todrive
/api/ttsfrom a real session.A silent path I noticed
If
data-speaking="1"ever sticks on the button, every later click is a silent no-op._playAudioBufonly clears the flag insrc.onended(success) or in thedecodeAudioDataerror callback. If the sharedAudioContextends up suspended,src.start(0)is inaudible andonendednever fires — the flag stays1and subsequentclicks do nothing and report nothing. Fits the symptom, but I can't confirm it from here.
Questions
missing — e.g. must the engine be (re)selected in Settings → TTS Engine in the UI,
rather than only being present in the server's persisted settings?
engine=elevenlabs?_playAudioBuf/speakMessagetreat a suspendedAudioContext(or a stalledstart) as a failure and resetdata-speaking, so a stuck flag can't swallow laterclicks?
All reactions