28 Commits
Author SHA1 Message Date
Dax 4d28a692d8 add fluid speech pro to the speech model catalog (#2007)
* add fluid speech pro to the speech model catalog

FluidVoice's Fluid Speech Pro, a Parakeet TDT 0.6B v3 retrain that also
declares Moondream's Parakeet Ultra as a base, for the same 25 European
languages. TalkMate converted it to a Q8_0 transcribe.cpp GGUF at d59dc79
(741,819,456 bytes, sha256 c3848e98...); the file says
parakeet/tdt-0.6b-fluid-speech-pro and announces its extra pre-Conformer
block as stt.parakeet.encoder.pre_block_kernel, so it needs a transcribe.cpp
that reads enc.pre_block.

Like Parakeet v3 it takes no language hint, detects the language under Auto,
does not stream, and is never recommended or badged. Its card publishes no
LibriSpeech test-clean figure, so it carries no single word error rate and no
benchmark record; its per-language FLEURS figures and the CC BY 4.0 credit
are in its details.

* bump(transcribe.cpp): 295c7e6, fluid speech pro's pre-conformer block

Fluid Speech Pro's GGUF announces stt.parakeet.encoder.pre_block_kernel and
carries eight enc.pre_block.* tensors: a residual FrameCleaner block between
the subsampler and Conformer layer 0. The old pin's graph never read them, so
the model loaded as plain Parakeet v3 and its encoder was wrong.

Both transcribe-cpp crates move to 295c7e601 on hex/combined-fluid, a
fast-forward of bcb639f that runs the block on pre_encode before xscaling
when the KV is present, and fails the load rather than skipping tensors when
it is only partly there. Cargo.lock changes only those two sources
(`getrandom`'s tempfile pointer was reverted to what it was), and
nix/package.nix carries the hash from a failed `nix build .#hex`, then a full
one whose check phase passed.

`hex model install fluid-speech-pro` followed by `hex model transcribe` on
whisper.cpp's jfk.wav prints the expected sentence, as does Parakeet v3 on
the same clip.

* fluid speech pro: snake_case id, FLEURS-only figures, blurb, notices
2026-11-19 15:19:06 -05:00
Dax 002b575c95 Add KB-Whisper Medium for Swedish (#1991) (#1992) 2026-11-19 05:44:56 -05:00
Kit Langton fb6c61b0d9 feat(cli): hex transcribe reads any common media file, with or without the app (#1799)
* Make hex transcribe work with any common media file, with or without the app (t1555)

* Regenerate the Labs goldens without Transcribe files (t1555)
2026-11-13 13:14:15 -08:00
Dax a38f6cb709 local api: native segment and word timestamps behind a labs toggle (#1658)
* local api: native segment and word timestamps behind a Labs toggle

POST /transcriptions takes timestamps=none|segment|word. Absent means
none and the response is unchanged. segment adds timestampGranularity
and segments; word adds segments and words, with text, startMs, and
endMs in whole milliseconds from the submitted audio's start.

The times are the model's own: the GGUF capabilities metadata decides
what it can produce, Apple Speech gives segments, and nothing estimates
or downgrades a level. A model without the requested level answers 422
unsupported-timestamps, decided after the model resolves on both the
shared dictation engine and the service's own copy. Chunked long clips
add each chunk's offset once from its sample position, and ends padded
past the decoded audio are clamped to durationMs.

Labs › Timestamps in API transcripts (labs.api_timestamps) gates it and
is off by default; off, a timed request answers 403 timestamps-disabled
before the audio body is read. The SDK takes timestamps and decodes the
new fields.

* local api timestamps: review follow-ups

* local api timestamps: correct embedded helper docs

* local api timestamps: align labs doc wording with the service spec
2026-11-10 17:07:52 -05:00
Dax 1843603caa local api: optional output processing behind a lab (#1597)
* local api: optional output processing behind a lab

* local api: share one processing state with listeners

* local api: process one transcript at a time

* docs: clarify the processing route checks
2026-11-09 10:35:05 -05:00
Dax 62a40669fe add orukeet to the speech model catalog (#1545)
* add orukeet to the model catalog

* orukeet: notice links, restored notice sentence, spec id, translations
2026-11-08 10:28:20 -05:00
Dax 8a327aa90d Add Granite Speech 5.0 470M TurboCTC (#1252) (#1256)
* Add Granite Speech 5.0 470M TurboCTC (#1252)

* fmt
2026-11-04 06:19:55 -05:00
Dax 637b1f933c Add NB-Whisper Medium for Norwegian (#1150) (#1173)
* Add NB-Whisper Medium for Norwegian, and list Norwegian (#1150)

* Embed q5_0 Vulkan shaders for NB-Whisper's file

* fmt
2026-11-03 12:43:26 -05:00
Dax efa1aa065a Add GigaAM v3 e2e RNN-T for Russian (#952) (#954) 2026-11-01 23:57:19 -05:00
Dax caa29177be Add Nemotron 3.5 ASR streaming 0.6B (#834) (#839)
* wip: Nemotron 3.5 ASR

* wip: Nemotron numbers

* Drop the temporary bench

* Docs for Nemotron 3.5 ASR

* Move the catalog's artifacts out of models.rs; fix tests for Nemotron
2026-11-01 09:48:16 -05:00
Dax 54d8f5387b Use speech models already on this computer (#594) (#819)
* Use speech models already on this computer (#594)

* Keep found-model state together

* CLI copy for found models

* Docs and onboarding hold test for models found on this computer
2026-11-01 07:24:24 -05:00
Dax dff0d7e005 Labs: Parakeet Ultra with a details view (#766)
* Labs: Parakeet Ultra catalog entry with details and attribution (#648)

* Tests, SDK id

* Say Parakeet Ultra needs its Lab

* Scene for Parakeet Ultra's details

* Docs, notices, and details never choose the model
2026-11-01 01:24:54 -04:00
Dax 5c5f0f66b4 Labs: use your own speech model (import a GGUF) (#759)
* wip: import your own speech model (Labs)

* fix debug

* Show the imported file's name

* clippy

* Say why an import can't be used

* Golden for the picker's Add model file

* Docs, feature map, AGENTS.md for imported speech models

* Arm Remove before removing an import
2026-11-01 01:03:12 -04:00
Dax d86d494c3b Add Hebrew with ivrit.ai's Whisper large-v3-turbo fine-tune (#755)
* Add ivrit.ai's Hebrew Whisper large-v3-turbo and Hebrew (#752)

* Move benchmark records into their own module

* Import ErrorMetric where the test uses it
2026-11-01 00:33:38 -04:00
Dax ee8de33bf2 Add QwenCleo-ASR 1.7B for Egyptian Arabic (#467) (#577) 2026-10-31 07:29:50 -04:00
Dax fdd6ccdc4a Add Underdog Bark 3.0 Pro and Clean transcription models (#345)
* feat: add Underdog Bark 3.0 Pro and Clean transcription models

Part of #343

* fix: use word_error_rate for Bark models

* fix: explicit wire names for Bark models
2026-10-29 01:53:35 -04:00
Dax 3434eb5f35 Rename voice-control to hex everywhere, without migrations (#293)
* Rename voice-control to hex everywhere, without migrations

The Cargo package and binary are now hex, every OS keeps its data in a
hex folder, and VOICE_CONTROL_* variables are HEX_*. Hex is a fresh
app: the macOS voice-control rename, its symlink, and the Windows and
Linux beta settings imports are gone. The Windows installer gets a new
AppId so it never takes over an old voice-control install's entry.

* Format

* Drop the Windows paste-last default only the beta import used
2026-10-28 13:08:44 -04:00
Kit Langton 4a5d659089 Local API: prepare never loads a second copy during dictation
When the dictation engine holds the requested model, prepare now uses it
and never loads its own. The engine publishes the GGUF selection it holds
after each command, so prepare can tell while dictation runs: a dictation
in progress means the weights are warm, and prepare succeeds at once
instead of waiting 20 s and then loading a duplicate onto Metal.

The prewarm is skipped when the session ran or streamed in the last
minute. The prepare stream waits on the result with recv_timeout instead
of a fixed 100 ms sleep, and reports the last stage before it ends. The
spec now says the prewarm runs over 1.5 s of silence, the minimum clip.
2026-10-24 16:30:04 -07:00
Kit Langton 96fbdc7517 Local API: preparing a loaded model reruns its prewarm
ggml keeps Metal weights resident for three minutes after a run; then macOS
compresses or swaps them and the next host clip pays to restore them
(+100-250 ms measured, more under heavy memory pressure). Preparing an
already loaded model, including the dictation engine's shared one, now runs
the prewarm over 1.5 s of silence, yielding to dictation like meeting chunks, so
a host can prepare when it starts recording.
2026-10-24 16:30:03 -07:00
Kit Langton f8774a5bbb fix(macos): stop meeting chunks at the shortcut press; share the dictation model with the local API
Meeting chunks now stop starting from the dictation shortcut press (including
a cold microphone open) until the clip is submitted, so a running chunk ends
within the hold threshold and release tail. The local API's host-audio
transcription runs clips up to 30 s on the dictation engine's model when it
matches, instead of loading a second copy (about 0.7 GB) on first use.
2026-10-22 02:17:43 -07:00
Kit Langton 9e9da53aa9 docs: audit guides against current behavior
Align desktop projections, processing fallback, history retention, contributor setup, Linux requirements, SDK contracts, and personal command examples with the implementation. Clarify prototype cleanup and historical validation limits without rewriting published release notes.
2026-09-01 09:56:40 -04:00
Kit Langton 7353e511d9 refactor: remove unused desktop and capture bookkeeping (#60)
Remove unused desktop projections, constant ingestion returns, and redundant recording-session state. Preserve capture behavior, shortcut recovery, and public SDK interfaces.
2026-09-01 09:45:51 -04:00
Kit Langton c1ddac1f9c docs(sdk): record embedding plan and 0.3 validation 2026-08-31 19:42:54 -04:00
Kit Langton 5b1d089612 fix: reject incompatible capture APIs 2026-08-06 20:09:39 -04:00
Kit Langton 62f06ac2d0 feat: default to Parakeet Unified English 2026-08-04 16:33:52 -04:00
Kit Langton 797b19c1c7 feat: prepare permanent macOS identity transition 2026-07-30 20:31:02 -04:00
Kit Langton c691f74941 feat: unify desktop modes and platform runtime 2026-07-23 17:33:22 -04:00
Kit Langton 3e166524bc feat: expand dictation tooling and TypeScript SDK 2026-07-22 20:58:27 -04:00