v0.8.0

speech, speech.cpp's one executable (tts, asr, voice, info, devices, models, pull, rm, quantize, serve and worker), and the shared library libspeech with its C API, speech.h, in one archive for each of macOS arm64 (Metal), Windows x64 (Vulkan) and Linux x64 (Vulkan, and the CPU alone; glibc 2.34 or later). The README lists the model families and where their converted weights are on Hugging Face.

What's Changed

• Give a recognition's result the languages Qwen3-ASR heard, as BCP 47 tags by @nyosegawa in #51
• Decode ReazonSpeech greedily when a request sets the option decoding by @nyosegawa in #52
• Let a Qwen3-TTS request set its sampling and the 1.7B model's instruction, and pass its audio from small chunks by @nyosegawa in #53
• Let an Irodori-TTS request and its file do what the official runtime does, and size the codec's windows by its speed by @nyosegawa in #56
• Install speech.cpp with one line, name models from a catalog and fetch them on first use, and try them on a page of speech serve by @nyosegawa in #57
• Shorten the README to an overview, move the details to pages in docs, and keep the workflows as skills by @nyosegawa in #58
• Quantize a model file in the library and speech quantize, each tensor as its family's layout table stores it by @nyosegawa in #59
• Keep the documentation to what the people who use speech.cpp need, and the rest in the code and the skills by @nyosegawa in #60
• Name the Irodori-TTS files of layout 2, which speak with the voice none and follow instructions by @nyosegawa in #61
• Transcribe a recording on the page while it is made, and send any audio in pieces so that an hour of it fits by @nyosegawa in #62
• Open the README with a banner of the project's name by @nyosegawa in #63
• Rewrite the decision records as the code decides today, merged by behavior and named without numbers by @nyosegawa in #64
• Keep in AGENTS.md what every task needs, and move releasing, model cards and the README's screenshot into the release skill by @nyosegawa in #65
• Give the README and the model pages the numbers measured with speech-bench for 0.8.0 by @nyosegawa in #66
• Say on Qwen3-TTS's page that the 0.6B model begins its speech with silence by @nyosegawa in #67
• Put the page's panels behind Speak, Transcribe and Live tabs by @nyosegawa in #68
• Detect where someone speaks with Silero VAD through speech_detect() and speech vad by @nyosegawa in #69
• Drop the worker's peek and raise its protocol to 3 by @nyosegawa in #70
• Detect speech in audio that arrives a piece at a time, with the regions speech_detect() gives by @nyosegawa in #71
• Transcribe a long recording by the regions where someone speaks, with speech asr --vad and chunking_strategy by @nyosegawa in #72
• Speak OpenAI's Realtime transcription over a WebSocket at /v1/realtime, with manual commits by @nyosegawa in #73
• Transcribe on the page by the regions a detection model finds, and whole only up to a minute without one by @nyosegawa in #74
• Open the page without a token, guarded by the Origin and the Host alone by @nyosegawa in #76
• Refuse an Irodori-TTS text whose predicted speech passes 30 s instead of squeezing it into 30 s by @nyosegawa in #77
• Speak a text a sentence at a time in speech tts and speech serve for a model whose request speaks less than a minute by @nyosegawa in #78
• Transcribe speech as it is said: speech asr --vad --live and -, and server_vad on /v1/realtime by @nyosegawa in #75
• Transcribe on the page's Live tab over /v1/realtime, with the detection model finding each utterance by @nyosegawa in #79
• Move the files into folders that say what each one is, and name the Qwen3-TTS checks after their family by @nyosegawa in #80
• Name the models to start with for speech asr given a lone WAVE file, and let two smoke scripts run on Windows by @nyosegawa in #81
• Open the README with a centered tagline, links and badges, and four lines on what speech.cpp gives by @nyosegawa in #82
• Release 0.8.0, with the page's three tabs near the top of the README by @nyosegawa in #83

Full Changelog: v0.7.1...v0.8.0