Repository navigation
CopySpeak v0.2.12
Added
-
Model picker for each engine — the model field for OpenAI, ElevenLabs, Cartesia and Google is now a list built from each vendor's documented models. A model id already saved in your config that isn't on the list stays selectable, so no setting is lost.
-
Tray menu follows the app language — the tray labels (the listening/paused toggle, Speak Clipboard, Settings, Quit) come from the same locale files as the window. Changing the language re-labels the tray without a restart.
-
Cyberpunk effect — a short synthesized 56k modem handshake opens and closes each reading, the first word and a few random later words glitch-stutter, and the voice gets a bright, lightly ring-modulated tone. Stutters add a little length, so karaoke captions drift slightly behind, as with the walkie-talkie lead-in.
Changed
-
OpenAI streams only on the
gpt-4o-mini-ttsfamily — the docs show chunked streaming for that model only. The defaulttts-1,tts-1-hdand unknown models now use the batch path, so default OpenAI users get their first audio only after the whole reading is synthesized. The docs don't confirm whethertts-1can stream. To stream, switch togpt-4o-mini-tts. -
Local engines stream only when warm — kokoro and pocket always play a reading as one piece. piper, kitten and qwen stream only once their daemon is running; until then a reading takes the batch path.
-
More of the interface is translated — about 45 more strings now go through the translation catalog. Arabic, German, French, Hindi, Russian and Chinese have drafts for every key, and Spanish and Hebrew are extended to match. The drafts are unreviewed; reviewers check them in Crowdin.
-
ElevenLabs defaults to
eleven_v4_turbo— fresh installs and the bundled ElevenLabs profile use the new model. A config that already saveseleven_v4keeps it; change the model in the ElevenLabs engine settings.
Fixed
-
Failed streamed readings finish cleanly — a reading whose stream errored mid-way could leave playback stalled, with other audio ducked until the next reading. The failed reading now stops and finishes, and ducking restores (#70).
-
Aborts no longer reach the next reading — an abort applies only to readings already running. A late error or leftover audio from an aborted stream can't affect a reading that started after it.
-
ElevenLabs
eleven_v4_turbokeeps its captions — the model sends audio-only records with no characters ahead of its text, and those cleared captions for the whole reading, so the live highlight never started. Those records now stay audio-only. -
Russian and other Cyrillic text is no longer sent to the engine with some letters swapped for Latin lookalikes ("привет" became "пpивeт"). Only words that mix Cyrillic and Latin letters, a common AI-text artifact, are still repaired.
-
Persian and Indic text keeps its zero-width non-joiner, so words spelled with it are no longer fused.
-
Hindi, Arabic, Urdu, Armenian, Amharic, Burmese and Khmer sentence endings now split long text between sentences instead of mid-word, and start a new caption phrase.
-
Playback events no longer fire twice — a page torn down before its event listeners finished registering left a listener behind, so every later event ran twice and a double-copy could speak twice.