README.md

---
title: audio.cpp ZeroGPU
emoji: 🔊
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.29.1
python_version: "3.10"
app_file: app.py
pinned: false
license: apache-2.0
short_description: 100+ TTS, ASR, cloning and music models via audio.cpp
tags:
  - audio
  - text-to-speech
  - speech-recognition
  - voice-cloning
  - ggml
  - gguf
  - zerogpu
---

# audio.cpp on ZeroGPU

A Gradio interface for [audio.cpp](https://github.com/0xShug0/audio.cpp), the native C++/ggml inference engine
for 100+ audio models: text to speech, voice cloning and voice design, speech recognition, forced alignment,
diarization, VAD, stem separation, voice conversion, restoration, and music or sound generation.

**How it runs.** On start-up the Space downloads the official audio.cpp CUDA 12.8 Linux prebuilt (the build used by
the project's Colab notebook) and NVIDIA's CUDA 12 runtime wheels. Each request starts `audiocpp_server` with a
one-model config inside a ZeroGPU call, posts to `/v1/tasks/run`, and stops it. Without a GPU the same binary runs on
CPU through a tiny `libcuda` stub (`stub/cuda_stub.c`).

**Inputs.** Type or paste text, load it from txt / md / srt / vtt / pdf / docx / epub / html / json files or a URL,
upload or record audio, drop any audio or video file (ffmpeg converts it), or pass audio URLs. Batch mode turns each
paragraph or line into its own clip, and audio tasks accept many files at once.

**Python tab.** Run Python with the `audiocpp_runtime` helpers, on ZeroGPU or CPU. Set the Space variable
`ALLOW_PYTHON=0` to disable it.

Environment variables: `AUDIOCPP_VERSION` (release tag, default `v0.9.0`), `MAX_GPU_SECONDS`, `MAX_BATCH`,
`AUDIOCPP_MODELS_DIR`, `ALLOW_PYTHON`.

Credits: audio.cpp by 0xShug0 and contributors (Apache-2.0). `model_params.json` comes from the project's WebUI
config. Model weights keep their original licenses.