audio.cpp
Local text-to-speech, speech-to-text and music without Python, run by audio.cpp on a CUDA GPU. Kokoro, Chatterbox, Fish Audio S2, OmniVoice, VibeVoice, BreezeTTS 2, Parakeet, Nemotron, ACE-Step and more.
The audio.cpp module runs GGUF audio models through audio.cpp's own server (audiocpp_server). It needs no Python and runs one process per loaded model, so a crash ends that server and not Voxta. One module has three services, each its own card in Add Modules:
| Service | Does |
|---|---|
| audio.cpp Text to Speech | Voices: built-in, cloned from a sample, or designed from a description. |
| audio.cpp Speech to Text | Speech transcription and clip transcription. |
| audio.cpp Music | Songs and instrumentals by ACE-Step 1.5. |
Experimental and admin-only. CUDA only, on Windows and Linux x64: Voxta downloads the server and its CUDA runtime on demand.
Setup
Add the service
Add Modules → Services → audio.cpp Text to Speech (or audio.cpp Speech to Text, audio.cpp Music) → Install, then follow the install steps.
Pick a model
In the install flow's Models step. Each service loads one model. You can also type a Hugging Face reference to any audio.cpp GGUF, or drop one in the models directory. Model Family is detected from the file. Set it only for a model audio.cpp cannot identify.
GPU Index picks the card when there is more than one. CPU Threads sets threads for the parts that run on the CPU.
Text to speech
| Model | Voices | Streams | Video memory |
|---|---|---|---|
| BreezeTTS 2 (default) | Clones a sample | Yes | ~4.5 GB |
| Kokoro 82M | Built-in | Yes | ~1.6 GB |
| Kitten TTS 2 | Built-in (Bella, Bruno) | No | ~3.5 GB |
| Chatterbox Turbo | One voice of its own, refuses a sample | No | ~0.9 GB |
| Chatterbox Multilingual | Clones a sample, needs one | No | ~1.8 GB |
| Fish Audio S2 Pro | Clones a sample | Yes | ~7 GB |
| Fish Drama 3 | Clones a sample | Yes | ~18 GB |
| OmniVoice | Clones a sample | Yes | ~0.6 GB |
| VibeVoice 1.5B / 7B | Clones a sample | Yes | ~2.8 / ~11 GB |
| Echo-TTS | Clones a sample, needs one | No | ~7.5 GB |
A streaming model's audio is sent as it is generated. The others send each line when it is done. Models that clone use the voice library and the character's voice sample. Several also use the sample's transcript. Default Female / Male Voice apply to the models with voices of their own.
Voice design
With BreezeTTS 2 or OmniVoice loaded, the service can design a voice in the Voice Studio. BreezeTTS 2 reads a description as written. OmniVoice takes a pick from its lists: gender, age, pitch, whisper style and accent.
Delivery
How a line is told its emotion tag depends on the model:
| Model | Delivery |
|---|---|
| Cloning models | Affect clips recorded beside the voice file. |
| Fish Audio S2, Fish Drama 3 | Clips, plus a bracketed tag in front of the line: [sad]. |
| BreezeTTS 2 | Clips, plus the direction as an instruction. |
| Kokoro, Kitten TTS 2, Chatterbox Turbo | None. |
The offered words are Delivery Directions (advanced), in the same shape as Fish Audio's audio tags: sad, laughing=decay(excited), moaning=repeat([moaning, panting, breathy], 4..9). Allow Unrestricted Directions (on by default) lets the AI write its own, like through gritted teeth, which Fish and BreezeTTS read as free-form directions.
Speech to text
| Model | Covers | Video memory |
|---|---|---|
| Parakeet TDT 0.6B v3 (default) | 25 European languages | ~0.8 GB |
| Fun-ASR-Nano 2512 | Chinese, Japanese and more | ~1.3 GB |
| Qwen3-ASR 0.6B / 1.7B | Chinese, Japanese and more | ~1.4 / ~2.8 GB |
| Nemotron 3.5 ASR Streaming 0.6B | Gives each word as it hears it | ~1.0 GB |
The other models re-read what you have said so far every Partial Interval (ms) (default 500, 0 turns partials off) and read it whole when you pause. Nemotron streams its words instead, and its final transcript is still read whole so it is punctuated. Max Segment Length (s) (default 30) finishes long stretches of speech without waiting for a pause.
Music
ACE-Step 1.5 in one GGUF file holding the music model, the planner, the text encoder and the audio decoder.
| Model | Notes |
|---|---|
| ACE-Step 1.5 Turbo (Q8_0) (default) | 8 steps. |
| ACE-Step 1.5 Turbo (BF16) | |
| ACE-Step 1.5 Base (Q8_0) | 30 to 50 steps, more detail. |
About 8 GB of video memory to make a song, 13 GB to repaint or remix one. It does Generate, Repaint and Remix. Takes, up to 4, are made one after the other. Length 10–600 s (default 120), BPM, key, beats per bar and a negative style are supported. There are no stems or LoRAs. Planner Thinking (on by default) lets the planner reason about structure, tempo and key first. For the Python version, with presets and LoRAs, see ACE-Step.