Speech Studio 0.5.0

Added

• Audio8 TTS Realistic-Air — a compact multilingual voice with zero-shot cloning (Settings → Models → Voicing): a new voicing engine beside Higgs and Supertonic — the Audio8 Realistic-Air 0.6B model (Q8_0, 1.5 GB) on the audio.cpp server. It speaks 23 languages, Russian among them, with its own stress, clones a voice from a reference clip and its exact transcript, and answers a phrase at a time at 44.1 kHz. Download it in Models; Higgs stays the default.
• Roles in Live: a trainer for any conversation — pick who the agent plays (Live → Role): an interviewer, an investor, a client after a discount, a sceptic, a boss who refuses a raise, a difficult colleague, an audience with a question, a language partner, a debate opponent. A role is a ready idea, not a frame: the agent may change it, and you two can make up your own just by talking — a hard conversation, a diction or language drill, an improvisation, a scene of a play — and keep it for next time (live_role_list, live_role_set). Every role is a file in the studio's roles folder that you can edit or add to. After the conversation the agent's review (live_role_review) gives your pace, pauses, filler words and how soon you answered, three of your phrases, three things to change and one goal, compared with your last time in the same role.
• Live: the end of your phrase heard by a model (Live → Microphone and sound → Agent's voice → «The end of your phrase»): Smart Turn v3.2 (32 MB, downloaded when you choose it) listens whether a phrase sounds finished, so a pause for breath cuts you off mid-sentence less often — at the cost of waiting a little longer for some ends. By the pause, as before, unless you choose it.
• The director inside an open project (#29): «Continue the script» writes what comes next for the same speakers, from the end of what is written, with an optional wish where it goes — the new blocks are added after the last one; «Rewrite the selected» words the chosen blocks anew as you ask and shows each before and after to keep or drop. Undo takes either back; agents have project_extend and project_enrich with mode rewrite.
• Quiet voices are no longer quiet (#34): a recording from the microphone is levelled as soon as it stops, a voice shows the loudness of its reference with a «Level the loudness» button for a quiet one (the original is kept), and in Live and while listening as it is voiced a quiet voice plays as loud as the others.
• Speech metrics for the agent in Live — voice_metrics gives the person's words a minute, pauses of 500 and 1000 ms, filler words (Russian and English, more in fillers.json), how soon they answer and their speech and silence; voice_lines gives their phrases with each word's time; a Live phrase carries word timings when the recogniser gives them.
• The agent sees when its speech stalls — say_state gives for each phrase the silence heard before it because it was not ready in time (queue_wait_ms) and how many such pauses were over 300 ms.
• A benchmark of Live for the agent — live_bench measures and keeps each run as a JSON file in the studio's bench folder, to compare two attempts: the waits of the conversation (median and 95th percentile), the rule that tells an interruption from a «mhm» on a marked set of phrases, and the word error rate of the recogniser on your recordings with their texts. The same from the command line: cargo run -p speech-server --example live-bench.
• Echo protection in Live — the agent no longer hears itself from the speakers: by default the microphone is not heard while the studio's sound is far louder than it, and you speaking over it are heard at once; or by your voice (voice gate), or by words only.

Changed

• Voicing no longer starts into a known failure: with a cloud engine that has no key or model, an engine that is not downloaded, or a graphics card the engine cannot run on, the window says so before anything starts and opens the settings that mend it; a block of pauses alone is skipped instead of failing; a voice pack download that broke off goes on from where it stopped.
• Higgs on a slow card is no longer cut off as hung: a phrase waits for the engine as long as that engine takes on this computer, the server is started again once when it ends before a phrase's first sound, and a clone from a short reference or one at 44.1/48 kHz is no longer refused by the codec.
• An update of a component is shown as an update — when a new version replaces an engine you already have, the start screen says «Components update» and names what is updated and why, instead of «First start».
• In Live the agent decides when you speak over it — the studio turns its voice down and the agent stops or goes on (voice_floor, say_continue); with no word from it within 1.2 s the studio stops by itself. The setting "When you speak over the agent" in Live's voice tab chooses the agent, an immediate stop or none. A phrase cut off while it played is shown as interrupted.

Fixed

• A block you approved by ear is exported (#35): a take the check turned down kept its block out of the export even after you listened to it and marked it «approved» — the export said hundreds of voiced blocks were errors. Your approval now outranks the check: the block counts as voiced, passes in the quality light and is exported as it is; the export window says so. The «Export without the unvoiced» checkbox shows its number again.
• A Russian take is checked in Russian: in a project with the language «Auto» the check guessed the language from the sound, and a Russian phrase in a voice with an accent was heard as Ukrainian and failed («сборки → збурки»). The check now hears a take in the language its text is read in. A new project on «Auto» also gets a library voice of its text's language instead of the first by name.
• Downloads through a proxy whose password has special characters (/ ? # :, a space, non-Latin letters) now work instead of being refused; the libraries the studio is built on are updated.
• Recognition no longer fails after voicing with Qwen3-TTS (#28): «transcribe.dll: LoadLibraryExW failed» came when the recognition engine was loaded after Qwen3-TTS — both carried a ggml library of the same name, and Windows gave the second engine the first one's. The recognition engine (transcribe.cpp 0.2.1) now names its own, so the two no longer meet in either order, on Windows and Linux.
• Blocks no longer stay unvoiced or keep a project from export because of a too strict sound check: a take was called «clipped» for a single run of three samples at full scale — now from 1 % of the samples, the level speech-quality tools fail a clip at — and such a take kept the whole project from export («blocks keep the export back»). A click at the end of a take is faded out instead of the take being thrown away. A phrase counts as «drawn out» only when it is twice as slow as normal reading (slow voices were turned away on every attempt), and the limits follow the speed you asked for. When every attempt still comes out with a defect, the least damaged take is kept and marked for you to listen to, instead of «No block was voiced». Every take turned away is written to the log with what it was measured by.
• The check of a take no longer mistakes a fast start or spoken numbers for a defect: an engine that begins speaking at once (Audio8) had nearly every take sent to a second recognition as «cut at the edge», which made checking a project take minutes; and a phrase with numbers said in words («восемнадцать тридцать») was called «mush instead of the text» when the recogniser wrote them in digits.
• A book with scene breaks voices to the end and a book of many hours exports (#24): a line of «* * *» became a block with nothing to say, so the count stopped short of the end (3165 of 3196) — it is now a pause after the block before it, and such blocks in projects already made are skipped. Exporting a whole book ran out of memory (an hour of sound took 8 GB): MP3 and FLAC, the loudness, the true peak and the limiter now work piece by piece.
• A book of thousands of blocks no longer freezes the window (#31): a project from 200 blocks draws only the blocks near the screen, the rest keep their place and are drawn as you scroll, search or jump to them. On 5200 blocks a click and a typed letter answer within a frame (the page did not answer for over 45 s before); dragging a block and the keyboard reorder work as before.
• A take sounds in the window as loud as its exported file (#27): the window played the model's own take, which with some voices and engines is 10–17 dB quieter than the export brought to the project's loudness — «quiet in the app, normal on disk». The player, the take history and the blind comparison now play a take at the project's loudness preset; the take's file stays as the model made it.
• An updated engine leaves no files of its earlier version behind: a download unpacked over the version before it removes that version's files the new one no longer has.
• Voicing blocks that have no voice no longer fails as a job: with an engine that speaks only a library voice the window opens the choice of a voice at once, and says to add one when the library is empty; a new project on such an engine starts with a library voice, and a voice of another language or without its reference text is no longer passed over.
• Russian speech the language detector takes for Belarusian is recognised in Russian by Parakeet, without asking to download CrisperWhisper.
• Live: the agent can be spoken over in a quiet voice — the echo protection measures how loud the studio's own voice comes back into the microphone (the window's volume and the microphone's gain are in it) and hears the person 12 dB over it; a muted microphone and a phrase cut by a stop are no longer taken for echo; a phrase that failed is shown as failed, not as sounding or interrupted.
• A take kept with a defect after every attempt is marked for listening whatever checks the project has on, and a failed export no longer erases the file exported before under the same name.
• Stresses, the voice gate and noise removal work on a computer without working CUDA: they take the processor build of ONNX Runtime there; when the runtime does not load, the error names its file and says to install the current Visual C++ 2015–2022 Redistributable.
• A project of one short word exports: «The sound is quiet: nothing to bring to the loudness» came for a programme shorter than 0.4 s.
• The voice gate recognises its owner again — your voice print was taken together with the silence around the recording and turned most of your phrases away; it is now taken from your speech only. Record your voice again in Settings, voice gate: an old print is not used until you do.
• Live no longer loses what the person says — the agent gets every phrase heard (heard is the number of lines there are to get); the studio's own voice heard back through the microphone is marked as echo, told by when its phrase sounded rather than by shared words, and a stop pressed in the window ends with the conversation.
• Qwen3-TTS downloads whole from the Models screen: the engine's build for the video card is shown beside the engine, so the engine no longer stays "not fully downloaded" after its files are there.
• The first start on a Mac no longer stops with an error before the recognition runtime is downloaded.