v0.8.8

speech, speech.cpp's one executable (tts, asr, vad, voice, info, devices, models, pull, rm, quantize, update, 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

• Pick Qwen3-ASR's tokens on the device, on a ggml argmax that takes the first of equal values with more threads by @nyosegawa in #155
• Give live transcription's readings the reading before as a draft, which Qwen3-ASR checks several tokens at a time by @nyosegawa in #156
• Run Metal's matrix products on the M5's matrix units without relaxed precision by @nyosegawa in #157
• Dequantize Q4_0, Q4_1 and Q4_K on Metal with their scales divided by 16 in float by @nyosegawa in #158
• Convert Drama 3 to one GGUF file and port its codec, checked stage by stage against the official one by @nyosegawa in #159
• Port Drama 3's audio decoder and its sampler, checked against the official ones by @nyosegawa in #160
• Port Drama 3's text model, checked a layer at a time against the official one by @nyosegawa in #161
• Write Drama 3's prompt and cut its long texts into chunks as fish-speech does by @nyosegawa in #162
• Speak with Drama 3 through the C API: its synthesis loop, its engine and its line in the table of families by @nyosegawa in #163
• Speak Drama 3 in the voices of voice files made of recordings and their transcripts, each given to a speaker by @nyosegawa in #164
• Move ggml to the fork's fix of Vulkan's batch stride of a view of the first rows of each batch by @nyosegawa in #165
• Compute on the CPU the experts of the layers a GPU cannot hold, chosen from its free memory when Drama 3 loads by @nyosegawa in #166
• Run a Drama 3 voice's recordings once and keep them for its next chunks and requests by @nyosegawa in #167
• Hold the model once at a time in the server smoke, and check Drama 3's options there by @nyosegawa in #168
• Name Fish Audio Drama 3 drama-3 in the catalog, fetching its Q4_K file by @nyosegawa in #169
• Release 0.8.8, with the numbers measured again on the Apple M5 by @nyosegawa in #170

Full Changelog: v0.8.7...v0.8.8