Aixora Audio

English · 中文

Your local audio studio
Speech synthesis · Speech recognition · Music and sound effects · Realtime voice · Video translation and dubbing

[!IMPORTANT] Aixora Audio 0.7.5 is available

Realtime voice calls from your phone, more supported models, an updated audio backend and bug fixes.

View the 0.7.5 release. Already using the 0.7.0 portable package? Run update.bat. For a first installation, download the complete 0.7.0 package, then run update.bat.

User guide · Build from source · Python setup · Model support · Acknowledgments

Aixora Audio brings local audio models into one workspace. Turn text into voiceovers, transcribe recordings into subtitles, translate and dub videos, or choose a character for a voice conversation. Audio models run locally through the C++ backend; conversation and translation can connect to a local or cloud language model you configure.

What you can do

Feature	Workflow
Speech synthesis and voice cloning	Enter text, choose a preset voice or reference recording, generate and listen; streaming models can play audio as it is generated
Speech recognition and subtitles	Transcribe recordings or videos, inspect timestamps, correct text and export subtitles
Video translation and dubbing	Transcribe, translate and review, generate speech per segment, then mix and export with optional background-music preservation
Realtime voice chat	Choose a character and voice, then talk on your computer or scan a QR code to call from your phone on the same Wi-Fi; automatic language mode follows the language you speak
Global dictation	Dictate into other applications using a shortcut, with recognition status shown in a floating toolbar
Music, sound effects and voice design	Generate audio from prompts or lyrics, describe a voice, or convert speech and singing voices
Audio tools	Separate vocals and accompaniment, audition stems, enhance and denoise, classify sounds, detect wake words, align text and convert formats
Model and reference-audio management	Download, load and remove models, save reference recordings with transcripts, and inspect devices and logs

Quick start

1. Extract the complete Windows portable package and double-click run_webui.bat in its root directory.
2. Once the browser opens at http://127.0.0.1:7899 by default, install the models you need in Model center.
3. Open a task page, select and load a model, enter text or upload your material, then generate.
4. Realtime voice chat and global dictation are under Scenarios. Configure the service URL, model name and API key in the relevant conversation or translation settings.
5. For phone calls, open System settings → Mobile access on your computer, enable mobile access, and scan the QR code with a phone on the same Wi-Fi network.

The portable package includes Python and runtime dependencies. The launcher selects an available backend in CUDA → Vulkan → CPU order; you can override it with AUDIOCPP_BACKEND. Languages, reference inputs and parameters are shown for the selected model.

To update the portable package, exit the workspace and double-click update.bat. For everyday use, these are the two entry points. Keep _env.bat and webui/start_webui.bat: the launcher calls them automatically.

0.7.0 is a new full-package baseline. Older packages can still update to 0.6.0 through the legacy channel; install the full Aixora-Studio.zip to move to 0.7.0. From 0.7.0 onward, updates use GitHub with Hugging Face and hf-mirror as fallbacks. Full packages use the same filename in the Hugging Face repository; incremental archives retain their versioned audiocpp-* names. See the updater guide.

When finished, choose System settings → Exit workspace service, or run run_webui.bat --stop. Closing the browser leaves the service running.

See the English guide / 中文使用说明 for detailed instructions. For a source checkout, follow the build and Python setup below.

Build the C++ backend from source

The main instructions below target Windows x64 with PowerShell. To run from source, build the C++ backend first, then set up the Python environment in the next section. The C++ backend performs audio-model inference; Python hosts the Aixora Audio workspace, model downloads, realtime helpers and SpeakType.

1. Prepare the tools and source

Tool	Requirement
Git	Used to clone the repository
Visual Studio Build Tools 2022 or later	Install Desktop development with C++, including MSVC x64, Windows SDK, CMake/Ninja and OpenMP support
CMake and Ninja	CMake 3.20 or later; the build script prefers the tools bundled with Visual Studio and also checks PATH
NVIDIA CUDA Toolkit (CUDA builds)	Install the full Toolkit with nvcc and an NVIDIA driver supporting your GPU; the driver alone is insufficient for compilation
Vulkan SDK (Vulkan builds)	Requires the SDK's glslc and a Vulkan-capable GPU driver; CPU builds need neither CUDA nor Vulkan SDK

git clone https://github.com/kigner/audio.cpp-webui.git audio.cpp
cd audio.cpp

Run the remaining commands from the repository root. A standard Aixora Audio build uses the vendored dependencies. The optional external/audio.cpp-server-frontends submodule is disabled by default, so a recursive submodule checkout is unnecessary for this setup.

2. Choose and build a backend

NVIDIA CUDA:

.\scripts\build_windows.ps1 -Preset windows-cuda-release -Target audiocpp_server -Jobs 16

CPU only:

.\scripts\build_windows.ps1 -Preset windows-cpu-release -Target audiocpp_server -Jobs 16

Vulkan:

.\scripts\build_windows.ps1 -Preset windows-vulkan-release -Target audiocpp_server -Jobs 16

Aixora Audio requires audiocpp_server for inference. Build the following optional tools for command-line inference or GGUF conversion. These examples use CUDA; replace -Preset for another backend:

.\scripts\build_windows.ps1 -Preset windows-cuda-release -Target audiocpp_cli -Jobs 16
.\scripts\build_windows.ps1 -Preset windows-cuda-release -Target audiocpp_gguf -Jobs 16

Outputs are written to build/<Preset>/bin/, for example build/windows-cuda-release/bin/audiocpp_server.exe. The source launcher discovers these directories. Keep the source tree intact, especially model_specs/, webui/configs/ and the model-manager scripts.

• -Jobs 16 controls build parallelism; use -Jobs 4 if memory is limited.
• CUDA builds detect the local GPU architecture through nvidia-smi by default. Use -CudaArchitectures to specify a target when building for other devices.
• -ConfigureOnly configures without compiling. -RunTests also builds and runs the C++ tests.
• The full model set is compiled by default, but backend compatibility varies by model.
• Aixora Audio downloads models through webui/model_manager_webui.py; it does not require -NativeModelManager. Model management for the upstream native WebUI is a separate feature.

3. Linux backend build (optional)

Install a C++17 compiler such as GCC 13+, CMake 3.20+ and Ninja/Make. From the repository root, build the CPU version with:

bash scripts/build_linux.sh --backend cpu --build-type Release --target audiocpp_server --target audiocpp_cli --target audiocpp_gguf

Outputs are written to build/linux-cpu-release/bin/. With the appropriate SDK installed, replace --backend cpu with --backend cuda or --backend vulkan. See the next section for running the Python workspace on Linux. SpeakType global dictation targets Windows.

Set up Python and launch Aixora Audio

Windows

Install 64-bit Python 3.11 (the project's development environment uses 3.11.9; the commands below use the Windows py launcher). After building the backend, run these commands from the repository root:

py -3.11 -m venv venv
.\venv\Scripts\python.exe -m pip install --upgrade pip
.\venv\Scripts\python.exe -m pip install -r webui\requirements.txt
.\venv\Scripts\python.exe -m pip check

webui/requirements.txt includes Python dependencies for Aixora Audio, realtime audio, SpeakType and the model-download tools. Windows desktop packages have platform markers. These commands use the virtual environment's interpreter directly, so Activate.ps1 is unnecessary. Installing Python packages with pip does not build the C++ backend.

Also install FFmpeg and FFprobe and add their executable directory to PATH, or place ffmpeg.exe and ffprobe.exe in webui/. They are used for non-WAV conversion, video input and dubbed-video export. A source checkout does not include the Python environment or these tools preinstalled in the portable package.

.\run_webui.bat

The launcher discovers the root venv. Open the URL printed in the startup log, normally http://127.0.0.1:7860; use the printed address if you configured a different port. Download a model in Model center, then load it on a task page. Model weights are not included in the source checkout.

To select a backend explicitly, set it before launching. Use cuda, vulkan or cpu, and build that backend first:

$env:AUDIOCPP_BACKEND = "cpu"
.\run_webui.bat

For a Python environment elsewhere, set AUDIOCPP_PYTHON to the full path of its python.exe. Aixora Audio starts its managed C++ service on port 8088 when needed; you do not need to start a separate backend process for the workspace.

Linux (optional)

Install Python 3.11 with venv support and FFmpeg/FFprobe. Microphone recording also needs the system PortAudio library. After building the backend, run from the repository root:

python3.11 -m venv venv
./venv/bin/python -m pip install --upgrade pip
./venv/bin/python -m pip install -r webui/requirements.txt
./venv/bin/python -m pip check
AUDIOCPP_BACKEND=cpu ./venv/bin/python webui/webui.py

Replace cpu in the last line if you built a CUDA or Vulkan backend. Open the workspace URL printed in the terminal and shut down the service from system settings when finished.

Setup troubleshooting

• PowerShell blocks the build script: run Set-ExecutionPolicy -Scope Process Bypass in the current window and retry. This setting lasts only for that PowerShell process.
• Build tools are missing: check the C++, Windows SDK, CMake/Ninja and OpenMP components in Build Tools; check nvcc for CUDA or glslc for Vulkan.
• Server not found: confirm that audiocpp_server was built and that AUDIOCPP_BACKEND matches your build.
• Python modules are missing: rerun dependency installation and pip check with the same venv; do not install into a different system Python.
• Video import or audio conversion fails: check that FFmpeg/FFprobe are on PATH or in webui/. For startup issues, read webui/logs/startup.log or run run_webui.bat --diagnostic.

Model support

Aixora Audio's Model center includes 129 model families and 184 model entries. The table below groups them by use. Select a model in the interface to see its versions, languages, reference inputs and parameters.

Category	Supported models
Speech synthesis and voice design	OmniVoice, Pocket TTS, DotTTS (SOAR / MeanFlow / Edit), NeuTTS 2E, Qwen3-TTS (0.6B / 1.7B) MioTTS 1.7B, VoxCPM2, VibeVoice 1.5B, IndexTTS2 / IndexTTS2.5, Irodori-TTS (v4 / v3) MOSS-TTS-Local v1.5, MOSS-TTS-Nano 100M, Supertonic 3, Higgs Audio v3 TTS 4B, Fish Audio S2 Pro GLM-TTS, Llama-OuteTTS 1.0 1B, VieNeu-TTS v3 Turbo, Inflect Micro v2, DramaBox Confucius4-TTS, Habibi / F5-TTS, MagpieTTS Multilingual 357M, Chatterbox, MOSS-VoiceGenerator Audio8 TTS Preview 0.6B, AuK, BreezeTTS 2, Chatterbox Turbo, CosyVoice3 Echo-TTS, FireRedTTS3, KittenTTS Mini 0.8, Kitten TTS 2, Kokoro 82M KugelAudio-0-Open, Maya1, MiraTTS, MOSS-TTS-v1.5, MOSS-TTSD Piper TTS, sanoTTS Nano, Soprano, Sopro V2 Turbo, VoxCPM1, ZipVoice Fish Audio Drama 3 / Drama3-small, Ming 3.0 TTS 0.5B, Ming 3.0 Audio Tiny, XTTS v2, Voxtral 4B TTS Plapre 3, LFM2.5-Audio (English / Japanese), FireRedAudio
Speech recognition and audio understanding	Qwen3-ASR (0.6B / 1.7B), Citrinet ASR, Nemotron 3.5 ASR 0.6B, Higgs Audio v3 STT, Hviske (v5.3 / v6) VibeVoice ASR, Voxtral Mini 4B Realtime, Fun-ASR-Nano 2512, Parakeet-TDT 0.6B v3, Kroko Community ASR SenseVoice-Small, Audio8-ASR-0.1B, Audio Flamingo 3 / Next, Canary 180M Flash, Cohere Transcribe Confucius4-R2T2, CrisperWhisper 2.0, FireRedAudio, GigaAM ASR, Granite Speech 5.0 470M TurboCTC Index-Echo-S2TT, LFM2.5-Audio (English / Japanese), Moonshine Streaming ASR, MOSS-Transcribe-Diarize, Niagara ASR OWSM, OWSM-CTC, SAMSONE, VibeVoice-ASR-BitNet, VibeVoice ASR Streaming Ming 3.0 Audio Tiny, Whistle ASR
Music and sound effects	ACE-Step 1.5, Stable Audio 3 (Small Music / Small SFX / Medium), HeartMuLa 3B, MiniMax-H3, MiniMax Music 3 ControlFoley, MiDashengLM-Gen, YuE2
Speech conversation	PersonaPlex, Ming 3.0 Audio Tiny
Voice conversion and audio codecs	MeanVC2, Vevo2, Seed-VC, RVC, MioCodec Tone Color VC
Source separation	HTDemucs, BS-RoFormer, Mel-Band RoFormer, HTDemucs 6-stem, SAM Audio, MossFormer2, TF-GridNet
Audio enhancement and restoration	Apollo, AudioSR, RNNoise, DeepFilterNet2, ZipEnhancer, GTCRN (Streaming / DNS3 / VCTK), FlashSR, RE-USE, Sidon v0.1 UniverSR
Voice activity and end-of-turn detection	Silero VAD, MarbleNet VAD, PulseVAD, FireRed VAD / Stream-VAD, Smart Turn v3.2
Sound classification	AST AudioSet, CED (Base / Mini / Small / Tiny)
Wake-word detection	microWakeWord, OpenWakeWord, Sherpa Zipformer KWS (Chinese / English)
Speaker diarization	Sortformer Diarization 4spk, Nemotron 3 Diarization, Sortformer Diarization v2.1
Subtitle and text alignment	Qwen3 Forced Aligner, MMS-300M Forced Aligner
Music transcription	MuScriptor Small, SheetSage2

Download models from Model center. Sound classification and wake-word detection are under Audio tools. MiraTTS and Sortformer v2 require manual weight installation.

Acknowledgments

Aixora Audio's audio inference backend is based on audio.cpp. Thanks to ShugoAI LLC and the upstream contributors. This project provides the Aixora Audio workspace, scenario applications, Windows launchers and portable-package updates.

See LICENSE for the code license.