RapidSpeech.cpp

Private speech AI that runs where your users are.

ASR · streaming ASR · TTS · VAD · voice cloning
One native C/C++ runtime. Python-simple. GGUF-powered.

Try in your browser · Run in Colab · Models · CLI guide

English | 简体中文

RapidSpeech.cpp gives applications local speech recognition, text-to-speech, voice activity detection and voice cloning without sending audio to a cloud service. The runtime is pure C/C++ on ggml, models use GGUF, and Python users get installable wheels instead of a heavyweight PyTorch stack.

Install, Then Speak or Transcribe

# Choose one package for your machine
pip install rapidspeech-metal  # macOS / Apple Silicon
pip install rapidspeech-cuda   # Linux / NVIDIA
pip install rapidspeech        # CPU, Linux / Windows

rapidspeech transcribe your-audio.wav             # speech → text
rapidspeech speak "你好，欢迎使用语音合成。" -o hello.wav   # text → speech
rapidspeech serve                                # OpenAI API + browser console

No model paths, no CMake, no config. The first run auto-downloads the default models (sha256-verified, resumable, cached in ~/.cache/rapidspeech/models/); inference uses no network after the files are cached. Set RAPIDSPEECH_OFFLINE=1 to enforce offline mode.

Example result from an Apple M1 Pro using Metal:

$ rapidspeech transcribe news-15s.wav
model: FunASRNano  backend: MTL0
15.10s audio in 2.08s (RTF 0.138)
国家发展改革委等部门发布关于开展重点行业节能降碳改造攻坚三年行动的通知。通知提出，重点行业能源消耗和二氧化碳排放的规模大、强度高。

Something broken? rapidspeech doctor checks your install, backends and model cache, and prints the exact fix for every problem it finds. See docs/cli.md for all commands (models, pull, offline mode, mirrors).

Want to see the whole loop at once? rapidspeech serve starts the native server, opens http://127.0.0.1:8080/webui.html, and exposes the browser console, OpenAI-compatible HTTP, and WebSocket streaming ASR/TTS endpoints from the same process. Models, backend, and per-request latency are shown on screen; nothing leaves localhost.

The default TTS model is Mandarin Kokoro v1.1-zh. Curated voice packs are available through the model registry, and English words embedded in Chinese text are supported. For English and voice-cloning TTS, see the model guides.

Try Before Installing

The browser demo runs inference client-side with WebAssembly and WebGPU:

• Hugging Face browser demo
• ModelScope browser demo
• Google Colab quickstart

It includes offline ASR, microphone ASR with VAD, and local TTS. Your model and audio stay in the browser session.

Use It From Python

rapidspeech pull funasr-nano     # prints the local model path

import os
import rapidspeech

model = os.path.expanduser("~/.cache/rapidspeech/models/funasr-nano-q4_k_m.gguf")
asr = rapidspeech.asr_offline(model)
asr.push_audio(pcm)              # 1-D float32 mono PCM
asr.process()
print(asr.get_text())

TTS, streaming ASR, VAD and voice cloning: Python examples.

What You Can Build

• Private transcription: offline files, VAD-segmented audio, hotword biasing for names and domain terms.
• Live captions and voice input: microphone streaming with partial and final results.
• Local speech output: multilingual TTS, expressive synthesis and zero-shot voice cloning.
• A local speech service: OpenAI-compatible HTTP endpoints, WebSocket streaming and MCP from one server.
• Browser and native applications: WASM/JavaScript, Python, C API and self-contained C++ executables.

Why RapidSpeech.cpp

1. Local by construction. Audio is processed by the native runtime on your machine. Model downloads are explicit and offline mode is enforceable.
2. A speech workflow, not only a model runner. VAD, streaming buffers, hotword biasing, Chinese text normalization, speaker embeddings, model quantization and serving live in the same project.
3. One core, several deployment surfaces. Use Python while prototyping, then embed the C API, ship a native executable, run a local server, or move the workflow into the browser.

The shared GGUF format targets CPU, Metal, CUDA, Vulkan, CANN, OpenCL and WebGPU. Backend support and optimization vary by model; the model guides record the current paths and known limitations.

Measured Performance

Apple M1 Pro, macOS, default models, warm inference, 5 runs each. The table is generated from raw JSON in benchmarks/results/ — see benchmarks/README.md for the rules and one-command reproduction on your own machine:

RTF is processing time divided by audio duration. Lower is faster; RTF below 1 means faster than real time. Results depend on hardware, backend, model, quantization and input; the CLI prints RTF on every run. CUDA / Windows / WASM platforms are marked "awaiting community reproduction" in benchmarks/README.md — we publish measurements, not estimates. The OmniVoice guide includes a per-model quantization/step matrix.

Supported Today

X-ASR — Chinese/English Zipformer2 transducer (icefall/k2). One GGUF serves both offline full-context decoding and true chunked streaming (per-layer left-context caches, sub-second partials, --chunk-len 16/32/48/96/192 fbank frames). Punctuation and casing, greedy transducer decode, runs on CPU / Metal / CUDA / Vulkan and quantizes to q4_k_m (99.5 MB).

IndexTTS-2 — expressive zero-shot voice-cloning TTS (GPT + S2Mel CFM + BigVGAN-v2 vocoder) with 4-mode emotion control (reference audio / vector / text / Qwen). See docs/index2tts.md.

Use It Your Way

In Progress

Qwen3-ASR, Qwen3-TTS.

Documentation

• Python examples
• Technical Notes: architecture, design tradeoffs, backends, model conversion, and binding surfaces.
• Model guides:
  • ASR — X-ASR (Zipformer2, streaming) · SenseVoice · FunASR-Nano
  • TTS — IndexTTS-2 (voice clone + emotion) · CosyVoice3 · OmniVoice · OpenVoice2 · Kokoro
  • VAD — Silero / FireRedVAD
  • Speaker — CAMPPlus
• Browser / WASM examples
• Node.js example

Native C++ CLI

Download Models

The pip CLI handles this automatically (see the top of this page). For the C++ CLIs, download GGUF files manually from:

• 🤗 Hugging Face: https://huggingface.co/RapidAI/RapidSpeech
• ModelScope: https://www.modelscope.cn/models/RapidAI/RapidSpeech

Build from Source

git clone https://github.com/RapidAI/RapidSpeech.cpp
cd RapidSpeech.cpp
git submodule sync && git submodule update --init --recursive
cmake -B build
cmake --build build --config Release

Self-contained executables (no runtime DLL/.so dependencies) — build the core and ggml statically into each CLI with -DRS_STATIC_EXE=ON:

# Windows / MSVC
cmake -B build -G "Visual Studio 17 2022" -A x64 -DRS_STATIC_EXE=ON
cmake --build build --config Release --parallel
# Linux / macOS
cmake -B build -DRS_STATIC_EXE=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

Build artifacts are located in the build/ directory:

• rs-asr-offline — Offline ASR command-line tool
• rs-asr-vad-online — VAD-segmented quasi-streaming ASR command-line tool
• rs-asr-online — True chunked streaming ASR (X-ASR; mic or WAV, live partials)
• rs-tts-offline — Offline TTS command-line tool
• rs-server — OpenAI-compatible HTTP API + MCP server (ASR + TTS)
• rs-quantize — Model quantization tool

Core Commands

Offline ASR

./build/rs-asr-offline \
  -m /path/to/funasr-nano-fp16.gguf \
  -w /path/to/audio.wav \
  -t 4 \
  --gpu true

VAD-segmented ASR

./build/rs-asr-offline \
  -m /path/to/funasr-nano-fp16.gguf \
  -v /path/to/silero_vad_v6.gguf \
  -w /path/to/audio.wav \
  -t 4 \
  --vad-threshold 0.5 \
  --silence-ms 600

Hotword biasing (FunASR-Nano)

# Bias proper nouns / fix homophones ; use --hotword-file for a large list
./build/rs-asr-offline \
  -m /path/to/funasr-nano-fp16.gguf \
  -w /path/to/audio.wav \
  --hotwords "阿里巴巴"

See docs/funasr-nano.md for details.

Streaming ASR (X-ASR)

# WAV, real-time paced with live partials (or --fast to run as fast as possible)
./build/rs-asr-online -m /path/to/xasr-q4_k_m.gguf -w /path/to/audio.wav --chunk-len 32
# Microphone
./build/rs-asr-online -m /path/to/xasr-q4_k_m.gguf --mic --chunk-len 16

See docs/x-asr.md for the model, chunk-size / latency tradeoffs, and GGUF conversion.

Text to speech

./build/rs-tts-offline \
  -m /path/to/omnivoice-f16.gguf \
  -t "Hello, welcome to RapidSpeech!" \
  --instruct "male, young adult, moderate pitch" \
  --lang English \
  --n-steps 32 \
  -o output.wav

Quantization

./build/rs-quantize /path/to/input-f16.gguf /path/to/output-q4_k.gguf q4_k

Server (OpenAI API + MCP)

# From the pip package: auto-download defaults, start API, open browser console
rapidspeech serve

# From a native build: serve ASR + TTS over OpenAI-compatible HTTP API and MCP
./build/rs-server --asr-model xasr.gguf --tts-model omnivoice.gguf --port 8080

curl http://127.0.0.1:8080/v1/audio/transcriptions -F file=@audio.wav -F model=rapidspeech-asr
curl http://127.0.0.1:8080/v1/audio/speech -H 'content-type: application/json' \
     -d '{"input":"hello","voice":"female"}' --output out.wav

Also runs as an MCP server (stdio for Claude Desktop, or POST /mcp) and exposes WebSocket streaming endpoints (streaming ASR with partial/final, VAD-segmented ASR, segmented + pure streaming TTS), plus a browser test console (--web-dir examples/server → http://host:port/webui.html). See examples/server/README.md.

Python

See Python examples for offline ASR, streaming ASR, offline TTS, streaming TTS, VAD, and voice cloning.

🤝 Contributing

Contributions are most useful when they include a command that reproduces the change and a result that can be checked. Current high-impact areas:

• Cross-platform installation and wheel testing.
• Reproducible latency, memory and quality benchmarks.
• Streaming, Metal, CUDA, Vulkan and WebGPU performance.
• Documentation and first-run examples.
• Model ports with conversion scripts, parity checks and known limitations.

Open an issue before starting a large port so the implementation can align with the runtime architecture.

Acknowledgements

1. Fun-ASR
2. llama.cpp
3. ggml
4. cppjieba — Chinese word segmentation
5. WeText — text normalization (ITN/TN)
6. miniaudio — single-file audio I/O
7. X-ASR Streaming-focused automatic speech recognition models