Model Gallery

6 models from 1 repositories

Filter by type:

Filter by tags:

localvqe-v1.1-1.3m
LocalVQE v1.1 (1.3 M parameters, F32) — joint acoustic echo cancellation, noise suppression, and dereverberation for 16 kHz mono speech. DeepVQE-style architecture with an S4D bottleneck and an in-graph DCT-II filterbank. ~9.6× realtime on a desktop CPU; 16 ms algorithmic latency. ~5 MB on disk. v1.1 ships the v16 echoaware checkpoint with improved double-talk and near-end single-talk AECMOS scores.

Repository: localaiLicense: apache-2.0

localvqe-v1.2-1.3m
LocalVQE v1.2 (1.3 M parameters, F32) — compact joint acoustic echo cancellation, noise suppression, and dereverberation for 16 kHz mono speech. Shares the same DeepVQE-style architecture (arch_version 3) as v1.3 but with narrower encoder/decoder widths, so it runs at ~9.7× realtime (~1.6 ms per 16 ms frame on a 4-thread Zen4 CPU) — about ¼ the per-hop cost of v1.3. Widens the echo-search window to 1024 ms (v1.1 used 512 ms). ~5 MB on disk. The budget-friendly choice for low-core or power-constrained devices.

Repository: localaiLicense: apache-2.0

localvqe-v1.3-4.8m
LocalVQE v1.3 (4.8 M parameters, F32) — current default release. Joint acoustic echo cancellation, noise suppression, and dereverberation for 16 kHz mono speech, with a wider encoder/decoder trained from scratch under a noise-floor-aware loss recipe. ~4.7× realtime (~3.3 ms per 16 ms frame on a 4-thread Zen4 CPU); ~19 MB on disk. Improves doubletalk speech quality (+0.25 deg MOS) and far-end echo cancellation (ERLE +5.2–9.3 dB) over v1.2; on far-end-only scenes some users may still prefer v1.2's gentler trade-off. Same 16 ms algorithmic latency as the compact models.

Repository: localaiLicense: apache-2.0

audio-cpp-htdemucs
HTDemucs (audio.cpp, F16) - four-stem music source separation (drums, bass, other, vocals), served by the audio-cpp backend through /audio/transform. One run produces all four stems; ask for a single one with the params[stem] form field, and read the X-Audio-Stems response header for the URLs of the rest. F16 rather than Q8_0 because upstream's tested-status table records Q8_0 as "Pass (drift)" while 16-bit is a clean Pass, and separation drift is audible in an isolated stem in a way it is not in a mix.

Repository: localaiLicense: mit

audio-cpp-seedvc-singing
Seed-VC (audio.cpp, Q8_0) - singing voice conversion, served by the audio-cpp backend through /audio/transform. Post the performance to convert as `audio` and the target singer's clip as `reference`. This entry pins task:svc, and the pin is what makes it a singing model rather than a speech one. Seed-VC advertises both svc and ordinary voice conversion, and nothing in a request says "this input is singing", so auto-routing always picks voice conversion; the pin is the only way to reach the singing path. Drop the option from the model config to get plain voice conversion from the same weights. Q8_0 because the 16-bit and Q8_0 builds carry the same tested status upstream (drift) and this one is half the size. Note the GPL-3.0 licence, which is stricter than most of the gallery.

Repository: localaiLicense: gpl-3.0

audio-cpp-vevo2-speech-to-speech
Vevo2 (audio.cpp, Q8_0) - speech to speech, served by the audio-cpp backend through /audio/transform. Post the source speech as `audio`, the target speaker's clip as `reference`, and the words being spoken as `params[text]`. The text is not optional and not a hint. Vevo2's speech-to-speech route is a text and prosody route: it reads the line it is resynthesising and refuses the run without one. /audio/transform has no text field on the wire, so the params map carries it and the backend lifts it into the engine's text input; `params[target_text]` is the same thing under the name Vevo2's own option table uses. Add `params[language]` when the line is not English. This entry pins task:s2s. Vevo2 also advertises plain voice conversion, which needs no text and which auto-routing picks first, so without the pin the text would be carried and then ignored. Drop the option to get voice conversion from the same weights. Q8_0 because it is the only build that fits alongside the rest of the gallery; upstream records the original and 16-bit builds as the cleaner ones, so expect drift on some routes.

Repository: localaiLicense: other