Speech Studio 0.2.4

Added

• The time of every step for agents (studio_timings): how long each phrase took to recognise and how soon after the end of your speech its text came, how soon after the agent's text its first sound played, and how long a voice's engine took to load, so the agent can compare recognisers and voices with you.
• The recogniser that hears is named: the top bar shows it next to Microphone on, and an agent asked to change it switches it at once, without restarting the microphone, while the Live window shows the new one.

Fixed

• Sounds no longer cut the agent off. A cough, a laugh or a sigh from Drama 3's tags, or a short "uh-huh" reaching the microphone from the speakers, was taken for you interrupting the agent; a phrase of at least two words now interrupts, as in Pipecat and LiveKit.
• The cross cuts speech off: ending the conversation also stops what the agent is saying.
• The voice an agent chose shows in the window: the Agent's voice panel follows the voice the agent set at your request.
• The installed app writes its log again; the diagnostics report showed "0 lines".

Speech Studio 0.2.3

Added

• The top bar shows while the microphone records: Microphone on, what it is on for (Live, a Listening recording, a voice recording, the microphone test) and for how long, with a cross that turns it off from any screen.
• Pause in Live (M): the agent falls silent and does not hear you; press again and the conversation goes on.

Changed

• Only the person turns the microphone on. Live, a Listening recording, a voice recording and the microphone test start only on your own click or key in the studio's window. An agent connected over MCP has no tool that turns the microphone on, and the window does not take its clicks on those buttons; the agent can only pause the conversation or end it.
• No window, no recording: when no window of the studio has been open for five seconds, the microphone turns off by itself.

Speech Studio 0.2.2

Added

• Live says the voice is loading. While the engine of a newly chosen voice loads, the window reads "Loading the voice…"; the engine loads at once on the change, not at the next phrase.
• Live warns about Parakeet and a chosen language. When a language is chosen and Parakeet recognises, the window says Parakeet detects the language itself and may write some words in another one, and suggests Nemotron 3.5, which keeps to the chosen language.

Fixed

• No strange pauses in Live. A long phrase of the agent is said whole up to 300 characters instead of being cut at a comma after 120, which gave Qwen3-TTS pauses and a new intonation mid-sentence; a pause an engine leaves inside a phrase lasts at most 0.6 s, as in an export with Even pauses.
• The aura has no square edge, and the wave follows the loudness smoothly instead of in steps.

Speech Studio 0.2.1

Changed

• The chosen recognition model is downloaded on first use. In Listening jobs, take checks and voices it is a step of the job with its progress; in Live it downloads in the background with progress in the top bar, and recognition starts by itself when it is done.
• A recogniser, a language or a microphone chosen during a conversation applies at once: the microphone starts again with the new choice.
• A new voice in a conversation loads its engine at once, not at the next phrase.
• Parakeet leaves video memory after a Listening job and a take check; in Live it stays while Live is on.
• Turning the microphone off answers the agent briefly: the recording's number and how many phrases were heard.

Fixed

• The agent's phrase is said once: a defective take is no longer voiced again aloud, and no check after the phrase voices it a second time.
• An agent turning the microphone on while it is already on in the window joins it instead of getting "Already recording"; the microphone test in the window gives the microphone up to the conversation.
• Two studio windows no longer take every connection: the main window and the widget each hold one event stream, so buttons no longer hang.
• A phrase the engine cannot say (a language it does not know, for one) is marked failed with its reason instead of waiting for ever.
• The bottom line names the model that actually recognises: its quantization and the graphics card or the processor.
• Nemotron 3.5 knows 28 languages; the window said 35.

Speech Studio 0.2.0

Added

• Live: talk to an agent by voice. An agent connected over MCP says each phrase as soon as it has written it; you answer into the microphone and are recognised phrase by phrase, and speaking over the agent stops it. The controls are those of voice modes: mute with M or Ctrl+Shift+Space from any window (hold to talk), text to the agent, the recogniser and the agent's voice, the agent's name on its lines, Ctrl+3 and the command palette. The Microphone and sound window has mixer-style level meters, Hear yourself, a speaker test and the agent's volume; five looks (aura, wave, bars, circle, matrix) and a widget over every other window.
• Drama 3 from 6 GB cards. When the model does not fit, its experts (18 GB) live in RAM (32 GB needed) and the card keeps a cache of the ones in use; the sound is the same. On an RTX 4090: whole ~8× faster than real time, 12 GB 4.7×, 8 GB 3×, 6 GB 2.2×.
• Voices of the pack cloned in Drama 3, and cloud voices played phrase by phrase while they voice.
• Gemini TTS with parallel requests or Batch at half price, and Game Dub's checks of Google's takes.
• Editing. What to fix on a take, Edit the phrase from the player, trimming a take on its waveform, A/B of takes, a blind comparison of engines, fixing a take by voicing again only the place of the error, approved blocks and notes on words.
• Text. An SSML subset, prosody hints coloured by word, phone numbers, e-mail and codes, a passage in another language in its own voice, the director placing punctuation and tags without changing a word, and a script check before voicing.
• Books and plays. Direct speech apart from the narrator's, a character register with voice casting, one voice per project, footnote cleaning, the book's names into the dictionary, a project updated from a corrected book without losing what was voiced, and an HTML book player.
• Montage. A track editor (shift, trim, fades, volume) with music and sound tracks from files, folders or a song made in YuE2, MiniMax Music3 or ACE-Step Studio.
• Export. SRT, VTT, ASS and LRC subtitles, sample rate and bit depth, voice effects (radio, phone, loudspeaker), even loudness of pieces, a file per chapter, a track per character in a ZIP, a CSV table of lines, name templates, a SHA-256 manifest and "how it was made".
• Voices. Recording from the voice picker, trimming a reference on its waveform, character variants, a voice by traits, a voice's licence and several pieces of a long reference.
• Quality. A traffic light and a project report, and a corpus of hard Russian phrases to check an engine.
• Nemotron 3.5 for recognition (28 languages with punctuation); in Live it recognises a phrase as it is spoken instead of again from its start.
• The recording's language under Auto is detected by VoxLingua107, and the recording goes to a recogniser that knows that language; speech is found by Silero VAD v6.2.
• Listening. Recording from the microphone with recognition on the go, a folder of recordings recognised in the background, quick recognition without the editor, a batch of files in several formats, subtitles by broadcasters' rules (Netflix, BBC and others), who speaks how much, a phrase repeat, numbers as digits in export, and the recogniser picked by the recording's language.
• The window and agents. Every route and action of the window is an MCP tool and the agent can be the studio's own language model; the Ctrl+K command palette, your own hotkeys, a simple mode, source tabs, Finished files, Restore after deleting a project, "?" hints on every element in six languages, verb-and-object buttons and a phone layout.
• The project site in six languages with samples of every engine: https://timoncool.github.io/speech-studio/

Changed

• Recognition on transcribe.cpp. Parakeet-TDT v3 and Parakeet Ultra moved from ONNX Runtime to transcribe.cpp, and the model stays loaded between jobs; on an NVIDIA card the CUDA build runs, without one the processor build, switched in the settings without a restart. Whisper stays on faster-whisper for the time of every word. The engine and the models are the ones Dub Studio uses, so either program takes them from the other. The ONNX recognition models of 0.1.0 are removed at start.
• First run asks only what you want: voices, recognition (one Whisper model or none) and extras; the engines, libraries and the build for your card come with them by themselves, named in one line with their size.

Fixed

• Voicing no longer stops on "stress lock" until a restart. When ONNX Runtime was downloaded after start (on the first-run screen), Windows loaded a foreign onnxruntime.dll 1.17 from System32, loading stress failed and every following block refused; the studio now always takes its own ONNX Runtime, and a failed load names its reason without stopping the next block.

Speech Studio 0.1.0

Added

• Five local voicing engines. Higgs Audio v3 (cloning, emotions, sounds), Fish Audio Drama 3 (for 24 GB cards), Supertonic 3 (ten voices, on the processor or the graphics card), Qwen3-TTS 1.7B (nine voices, cloning, a direction for each block) and Qwen3-TTS with Russian stress (a LoRA attached on the fly); cloud voices from OpenRouter, Google Gemini (with Batch at half price), fal.ai and Runware, each asked for first.
• Projects. Paragraphs become blocks and Anna: lines a dialogue; takes, stale marks, undo, drag and drop, search and replace with regular expressions, notes and markers; an engine and a voice of its own for every character and every block.
• Russian text as it is spoken. Numbers in the right case, stress, «ё» and homographs through RUAccent (its big_poetry model in int8, optional), a pronunciation dictionary with import from Balabolka and a check by ear, markup compatible with SpeechKit.
• Every take checked. A take is recognised back and compared word by word, checked for defects (clipping, silence, a cut-off, a click at the end, a repeated phrase, a tail, music or noise instead of speech) and for voice similarity; several takes keep the best one.
• Podcasts and books. The director writes a podcast for any number of voices and lays a book chapter out by characters; EPUB, FB2, PDF, DOCX, HTML, Markdown and TXT come in; M4B with chapters, parts for long books and an ACX check go out.
• Subtitles as a project. SRT, VTT and ASS lines are voiced at their own time, with the pace raised ahead, a retry, a stretch and silence before the line; the export timeline keeps, trims or joins overlapping lines.
• Export. WAV, FLAC, MP3, OGG, M4A and M4B, subtitles and synced text, a file per block, loudness presets with a true-peak limiter, room tone and breaths, and a check before export (no take, stale, failed, file changed, too long).
• A voice from a description with Qwen3-TTS VoiceDesign: variants to listen to, and the one you like goes into the library.
• Voicing heard as it is made. Higgs and Drama 3 play while they voice, with the time to their first sound.
• Listening. Recognition with Parakeet-TDT or Whisper (with a prompt), up to eight speakers, long recordings cut at pauses by Silero VAD v6, noise removed by DPDFNet and music by Roformer, several files at once; a transcript with speakers, word times, a karaoke player, search, cutting by the text, a hallucination filter and edits, exported to SRT, VTT, ASS, TXT and JSON; a voicing checked against its text, a known text laid over a recording, and voices taken from the speakers.
• Voice library. A voice pack, 743 Russian voices from the cloud catalogue, recording from the microphone, files and ZIP, a photo and a card, a reference's transcript recognised once, cleaning with Roformer or DPDFNet keeping the original, and the best piece of a reference.
• First run and the model manager. Resumable downloads checked by SHA-256, import from a folder, and models from the author's other studios taken without downloading them again.
• Agents. An MCP server with every feature and the window, the assistant panel and an OpenAI-compatible audio API.
• Everywhere. Diagnostics, a bug-report package without keys, six languages, light and dark themes, a question before closing while jobs run, auto-update, and anonymous statistics you can turn off.