v1.22.0

1.22.0 (2026-11-11)

Features

• serve: report a segment request the client dropped while it waited (#230) (08abebd)

CTC alignment models v1

CTC forced-alignment models for lyrics/karaoke alignment. Each model is an fp16 ONNX export of the upstream checkpoint, with a vocab.json (id to char, blank, delimiter, sampleRate, frameMs). Input input_values, output raw logits (20 ms frames). This release is immutable and not an engine release.

id	languages	source checkpoint	licence	model sha256	vocab sha256
wav2vec2-ctc-fr	French	https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-french/tree/7c79e105a6525d38e1e69f640b974b4a679723cc	Apache-2.0	b868c38eb5e124ffb6ddbcd8482f47239cb4066db97b67ceb47a9c80c2bd62db	061a54098e7993f516da7edb41d0404411ff9cb786c7bbcf2f69b1926a6bf060
wav2vec2-ctc-en	English	https://huggingface.co/facebook/wav2vec2-large-960h-lv60-self/tree/54074b1c16f4de6a5ad59affb4caa8f2ea03a119	Apache-2.0	cc20dc5db7ed39b8f3bbbb34bbb5a9b3a3e1468701ed414562901c134275cdfd	11c0856d5dbcfd9f3cf8a8d138a782df7e6d1e7f97e4315bc1e4b1609aeaa09e
omniasr-ctc-300m	Multilingual (used for de, es, it, nl, pt and any other language)	https://huggingface.co/facebook/omniASR-CTC-300M/tree/8e35f0cc28fa6099e0c14d56db85ce0423baa691 (omniASR_CTC_300M_v2; ONNX conversion by https://huggingface.co/OpenVoiceOS/omnilingual-asr-ctc-300m-onnx/tree/bc55e1de19af8a907b0a010d06fc4637a2d5bf0c)	Apache-2.0	1c92b3c113e9a3574b812e039433db2bf588d31ca90769179d2a4acf1b9fddd0	333eb690077bb195de573f85afbfb0b6ec296c520c8b416ec9fd7edaa0dee104

v1.21.3

1.21.3 (2026-11-09)

Bug Fixes

• decide: read video audio codecs only from video profile entries (#226) (4a58a12)

v1.21.2

1.21.2 (2026-11-09)

Bug Fixes

• item: run each part's encode to its end, throttled like a video run (#225) (5239e25)
• runtime: kill playback runs instead of stopping them, move the cache to v2 (#223) (fdd1c28)

v1.21.1

1.21.1 (2026-11-09)

Bug Fixes

• item: drop the segment a run's budget cuts off (#221) (871ab7d)

v1.21.0

1.21.0 (2026-11-09)

Features

• karaoke: run the Windows ML runtime of the machine, ship none of it (#219) (fb01cc1)

v1.20.0

1.20.0 (2026-11-08)

Features

• karaoke: run the recogniser on Vulkan, drop whisper base (#217) (1e525d5)

v1.19.1

1.19.1 (2026-11-08)

Bug Fixes

• karaoke: catch gpu failures that only show in a run, cache migraphx programs (#214) (8b32241)
• runtime: report the served rung, not the client's pick (#215) (74d6020)

v1.19.0

1.19.0 (2026-11-08)

Features

• karaoke: run the model jobs on Windows ML, DirectML or the CPU (#210) (e9d4191)

v1.18.0

1.18.0 (2026-11-08)

Features

• karaoke: ship the karaoke tools in the macOS and Windows bundles (#207) (c6c93e2)