Speech Studio 0.3.5

Added

• The interface can be scaled (#16): Settings → Appearance → Interface scale draws the whole studio from 90 to 150 % and keeps it for the next start; Ctrl + plus and Ctrl + minus, or Ctrl with the mouse wheel, change it too, and Ctrl + 0 returns to 100 %.
• The voice catalogue filters, searches and plays before you download: the extra voices come with their name, gender, age (adult or child), language, description, length and year of birth, so the library's "Open cloud voices" filters by gender, age, language and minimum length, searches the name and description, sorts by name, length or age, plays a sample, and downloads the ones picked or everything the filters show, with the sample's transcript where the collection has one. A downloaded voice keeps all of it: the library shows it under its own name with its description, length and year of birth. The catalogue holds 1,636 voices: 1,394 Russian and 242 in 31 other languages (English 75, French 28, Italian 16, German 14, Portuguese 13, Spanish 13, Polish, Arabic and Turkish 9 each, Japanese 7 and more); 1,068 male, 568 female, 60 children. If you know a good collection of popular voices in another language, open an issue and we will add it.
• The library finds voices as the catalogue does: the same search over names, descriptions and tags, filters by gender, age, language, source, length and favourites, sorting by name, length or age, a count of what is shown and «Reset filters»; a project's voice field finds a voice by its name or description.
• A voice without its reference text is heard when it is voiced: the reference's text is recognised once, before voicing, for every engine that clones, so a catalogue voice without a transcript (about half of them) or an imported one needs no «Recognise» first; Drama 3 no longer refuses such a voice.
• The API takes the engine's settings (#17): POST /v1/audio/speech and the MCP tool speak_text accept, beyond OpenAI's fields, the project settings of the local engines — seed, temperature, top_k, top_p, max_tokens, steps (Supertonic) and loudness (off, video, podcast, acx) — each over the engine's default, as Project settings set them in the window.

Fixed

• Every engine runs on any x86-64 processor, with or without AVX2, FMA or BMI2: Drama 3 starts again as it did before 0.3.0, which refused it with «Drama 3 needs a processor with AVX2…», and the audio.cpp engine (Higgs, Supertonic, CrisperWhisper, Parakeet), Qwen3-TTS, the voice separator and recognition (transcribe.cpp) now come with a processor build for every instruction level and load the one this processor runs best, instead of the AVX2 build alone. The new builds are downloaded once after the update.
• A voice reference in a language the recogniser does not know is heard anyway: under Auto, a Chinese (or other) reference recognised by the default Parakeet, which does not know it, came back empty and was refused as «no speech was recognised»; now the reference's language is heard by VoxLingua107 and a recogniser that knows it (CrisperWhisper) is used, so its transcript, its cleaning and its voicing with Drama 3 go through.
• Auto only guesses a language a recogniser can read: the language id (VoxLingua107) names a few languages CrisperWhisper has no token for (Esperanto, Guarani, Interlingua) and a recording guessed as one of them was refused; Auto now takes the likeliest language a recogniser here can read instead.
• The voice-similarity check runs on the processor when there is no usable graphics card, instead of stopping a voicing with an ONNX Runtime CUDA error: its model asks for the card only when one can run a CUDA build, as under Auto with no card.