Filter by type:
Filter by tags:
audio-cpp-yue2-3b
Yue2-3B is a text-to-music model that generates songs from lyrics and a
style prompt, with optional symbolic ABC melody or score conditioning. This
audio.cpp package uses the published Q8_0 main model and F16 VAE pairing and
preserves the model, generation, tokenizer, and VAE sidecars it requires.
The model and GGUF package are licensed CC-BY-NC-4.0 for non-commercial use.
Review the upstream license before installation or deployment.
Repository: localaiLicense: cc-by-nc-4.0
audio-cpp-fish-drama3
Fish Audio Drama 3 is an experimental community port served by audio-cpp.
This GGUF uses Q4_K_S experts and an 8-bit (Q8) backbone. The download is
approximately 24.6 GB; runtime memory requirements exceed the weight size.
Allow additional memory for inference and reference audio processing.
For voice cloning, supply a server-local reference WAV path in the OpenAI
`voice` field and its transcript in `params.reference_text` (or
`params.ref_text`). Cloning requires reference audio and a matching
transcript; plain TTS can run without a reference. No named preset voices
are provided. Use only reference audio for which you have rights and consent.
The Fish Audio Community License governs the weights, not MIT. Review its
use and redistribution terms before installation. This community port is
experimental and is not an official Fish Audio serving implementation.
Repository: localaiLicense: other
audio-cpp-indextts-2.5
IndexTTS 2.5 is a multilingual text-to-speech model with zero-shot voice
cloning and emotion control, served by the audio-cpp backend. Supply a
server-local WAV path in the OpenAI `voice` field as the speaker reference.
This entry uses the published original-dtype GGUF. The package preserves
the source tensors without another quantization step, and audio.cpp records
the original-dtype IndexTTS2 path as passing its validation suite.
The bilibili Model Use License governs the model. Review its use,
redistribution, and large-scale commercial deployment terms before
installation.
Repository: localaiLicense: other
audio-cpp-supertonic
Supertonic 3 (audio.cpp, GGUF orig) - fast English text to speech with no
reference clip required, served by the audio-cpp backend. Covers both the
TTS and the TTSStream RPC, so /v1/audio/speech works streaming and
non-streaming.
The orig (F32) package is deliberate. Upstream's tested-status table in
docs/gguf.md records supertonic as Pass on orig, untested on 16-bit, and
"No (unsupported weight dtype)" on q8_0; the f16 package aborted the
process inside ggml_concat when tested locally. The backend refuses
anything but f32/i64 weights for this family, so a wrong build fails at
load with a clear message rather than crashing.
Repository: localaiLicense: openrail
audio-cpp-higgs-audio-v3
Higgs Audio v3 TTS 4B (audio.cpp, Q8_0) - expressive conversational
speech in more than 100 languages, with zero-shot voice cloning and inline
control over emotion, style, prosody, pauses and sound effects.
Supply a server-local WAV path in the OpenAI `voice` field to clone a
speaker, and use only audio for which you have the necessary rights and
consent. The Q8_0 GGUF is the recommended quality and speed balance and is
recorded as passing upstream's audio.cpp validation.
The Boson Higgs TTS 3 Research and Non-Commercial License permits research
and non-commercial use. Its creator-use grant and commercial-use terms
carry additional conditions; review the model license before deployment.
Repository: localaiLicense: other
audio-cpp-chatterbox
Chatterbox (audio.cpp, Q8_0) - multilingual zero-shot voice cloning, served
by the audio-cpp backend.
A reference clip is REQUIRED. This build advertises only the cloning and
voice-conversion tasks, not plain TTS, so every request must set the TTS
voice field to the path of a WAV file on the server; a voice that is not a
file on disk is treated as a named preset and the request is refused with a
message saying so.
The voice-conversion half is reached through /audio/transform rather than
/v1/audio/speech: post the speech to convert as `audio` and the target
speaker's clip as `reference`. That is why this entry declares
audio_transform alongside tts. With both this and a source-separation model
installed, /audio/transform has two candidates, so name the model
explicitly rather than relying on the default pick.
Q8_0 rather than 16-bit because upstream's tested-status table records the
same result for both (Pass, ASR match with drift) and this build is half
the size.
Repository: localaiLicense: mit
audio-cpp-citrinet-asr
Citrinet ASR (audio.cpp, Q8_0) - compact English speech to text, served by
the audio-cpp backend through /v1/audio/transcriptions. 40 MB, offline
transcription only. Q8_0 is the only package upstream publishes and tests
for this family, and it is recorded as a clean Pass.
Licensing: the checkpoint is pulled from NVIDIA NGC
(nvidia/nemo/stt_en_citrinet_256), whose use is governed by the NGC Terms
of Use rather than a named open licence.
Repository: localaiLicense: other
audio-cpp-nemotron-asr
Nemotron 3.5 ASR Streaming 0.6B (audio.cpp, F16) - streaming English speech
to text, served by the audio-cpp backend. It serves the backend's live
bidirectional transcription RPC as well as ordinary
/v1/audio/transcriptions calls.
F16 rather than Q8_0 because upstream's tested-status table records Q8_0 as
"Pass (minor filler drift)" while 16-bit is a clean Pass, and drift in a
transcript is drift in the output a caller reads.
Repository: localaiLicense: other
audio-cpp-sortformer-diarization
Sortformer Diarization 4-speaker v1 (audio.cpp, Q8_0) - speaker diarization
for up to four speakers, served by the audio-cpp backend through
/v1/audio/diarization. Returns per-segment start, end and speaker label; it
does not transcribe, so pair it with an ASR model for text.
Q8_0 because upstream records it as a clean Pass, the same as 16-bit, at
two thirds of the size.
Licensing: the base checkpoint nvidia/diar_sortformer_4spk-v1 is
CC BY-NC 4.0, so commercial use is not permitted.
Repository: localaiLicense: cc-by-nc-4.0
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-stable-audio-sfx
Stable Audio 3 Small SFX (audio.cpp, Q8_0) - text to sound effects, served
by the audio-cpp backend through /v1/sound-generation. Set
duration_seconds to control the length.
Text prompt only. The family also has an audio-conditioned editing route
upstream, which is not wired here: it aborts the process on the pinned
revision, and no field on the sound-generation request can reach it in any
case.
Q8_0 because upstream records both 16-bit and Q8_0 as "Pass (drift)", so
the smaller build costs nothing extra.
Repository: localaiLicense: other
audio-cpp-forced-aligner
Qwen3 Forced Aligner 0.6B (audio.cpp, Q8_0) - aligns a transcript you
already have against the audio, served by the audio-cpp backend.
It is driven through /v1/audio/transcriptions, not a dedicated endpoint:
put the known transcript in the prompt field and the model returns it with
word timings instead of recognising anything. Both a transcript and a
language are required, so this entry defaults language to en; override it
with the language form field for other languages.
Q8_0 because upstream records both 16-bit and Q8_0 as a clean Pass.
Repository: localaiLicense: apache-2.0
audio-cpp-silero-vad
Silero VAD (audio.cpp) - voice activity detection over /v1/vad. The model
ships inside the audio-cpp backend package, so installing this entry
downloads nothing.
Repository: localaiLicense: mit
audio-cpp-marblenet-vad
MarbleNet VAD (audio.cpp) - NeMo voice activity detection over /v1/vad, an
alternative to Silero with different segment boundaries. The model ships
inside the audio-cpp backend package, so installing this entry downloads
nothing.
Repository: localaiLicense: other
audio-cpp-irodori-voicedesign
Irodori TTS 600M v3 VoiceDesign (audio.cpp, Q8_0) - describe the voice you
want in words instead of supplying a reference clip, served by the
audio-cpp backend.
Drive it through /v1/audio/speech with the `instructions` field: the
backend routes a request carrying instructions to audio.cpp's vdes task, so
"a calm elderly woman speaking slowly with a warm, gentle tone" shapes the
delivery of whatever you put in `input`. Without instructions this build
still answers as ordinary TTS, and with a reference WAV in `voice` it
clones instead; the request decides which of the three it gets.
The model is Japanese-first and handles English, so an English ASR round
trip recovers the words with accent drift rather than cleanly.
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