Open-source text-to-speech you run yourself

voicestudio leads with a score of 98. Updated from public GitHub data.

Full category: Voice.

Which one?

1st pickvoicestudio
Desktop app for voice cloning, dubbing, dictation and audiobooks
Score 98 · 113k stars · AGPL-3.0
Switch toPocket TTS
CPU-only text-to-speech model with voice cloning and streaming output
for a permissive license (MIT) · Score 76 · 10.3k stars · MIT
Switch towhisper.cpp
C/C++ speech-to-text engine running OpenAI Whisper models locally
for a permissive license (MIT) · Score 90 · 55.3k stars · MIT
Open-source text-to-speech you run yourself
RankProjectAdoptionFreshnessMaintenanceAgent-readyScore
1voicestudioDesktop app for voice cloning, dubbing, dictation and audiobooks113k starsAGPL-3.0last commit Nov 202699100998598 out of 100
2Pocket TTSCPU-only text-to-speech model with voice cloning and streaming output Live demo ↗ (opens in a new tab)10.3k starsMITlast commit Nov 202656100883076 out of 100
3whisper.cppC/C++ speech-to-text engine running OpenAI Whisper models locally55.3k starsMITlast commit Nov 202694100973090 out of 100

Momentum, Verified build, Docs and Privacy are not measured yet; their weight goes to the signals shown. A dash means the signal is not scored for that kind of project. Hover a number for its rating in words.

Reviews

198 out of 100

voicestudio

Desktop app for voice cloning, dubbing, dictation and audiobooks

113k starsAGPL-3.0last commit Nov 2026

VoiceStudio is an Electron desktop app with a Python backend for voice cloning, voice design from text descriptions, timed video dubbing, dictation, transcription and audiobook creation. The default engine is k2-fsa/OmniVoice, and other engines can be selected. It also exposes a local API and an MCP server for agents.

Strengths

  • Covers cloning, voice design, dubbing, dictation and audiobooks in one app
  • Runs on NVIDIA CUDA, Apple Silicon (MPS) or CPU only
  • Local API and MCP server for agent integration
  • Installers for macOS, Windows and Linux, plus Docker

Weaknesses

  • Model weights and voices carry licence terms separate from the AGPL-3.0 app
  • CPU-only use is slower; the CPU PyTorch build needs about 5 GB disk
  • Intel Mac runs the UI only and needs a remote backend
  • Windows on ARM support is experimental and CPU only
  • GPU optional
  • Docker + Compose
  • Needs Python, PyTorch, Electron
  • Models: k2-fsa/OmniVoice
276 out of 100

Pocket TTS

CPU-only text-to-speech model with voice cloning and streaming output

Live demo ↗ (opens in a new tab)10.3k starsMITlast commit Nov 2026

Pocket TTS is a roughly 100M-parameter text-to-speech model from Kyutai that runs on CPU using about 2 cores, with a Python API, a CLI and a local HTTP server with a web UI. It streams audio with about 200ms to first chunk, clones voices from a wav file, and supports English, French, German, Portuguese, Italian, Spanish and Dutch. Small English variants (18.2M, 6.9M, 1.9M parameters) and training code are also published.

Strengths

  • Runs on CPU with 2 cores; about 6x real time on a MacBook Air M4
  • Voice cloning from a wav file, with export to fast-loading safetensors
  • Standalone Rust binaries for Linux, macOS and Windows need no Python or PyTorch
  • Training code released, plus 1.9M to 18.2M parameter English models

Weaknesses

  • GPU is not officially supported; serve and the Docker image always run on CPU
  • No way to insert silence or pauses in the input text
  • Linux pip install pulls CUDA PyTorch (about 3 GB) unless the CPU index is set
  • Rust binary is CPU-only and cannot load training checkpoints
  • GPU optional
  • Docker + Compose
  • Needs PyTorch 2.5+, Python 3.10-3.14, Hugging Face (model and voice downloads)
  • Models: Pocket TTS 100M, english_18m_2026-10, english_7m_2026-10, english_2m_2026-10, community models via --config
  • port 8000
390 out of 100

whisper.cpp

C/C++ speech-to-text engine running OpenAI Whisper models locally

55.3k starsMITlast commit Nov 2026

whisper.cpp is a dependency-free C/C++ port of OpenAI's Whisper ASR model, built on ggml, with a C-style API in whisper.h and a whisper-cli example for transcribing 16-bit WAV files. It runs on CPU or on accelerators including Metal, Core ML, CUDA, Vulkan, ROCm, OpenVINO, Ascend CANN, MUSA and AMD Ryzen AI NPU, and supports integer quantization and VAD.

Strengths

  • No runtime dependencies; CPU-only inference works, with zero memory allocations at runtime
  • Broad backend coverage: Metal, Core ML, CUDA, Vulkan, ROCm, OpenVINO, CANN, MUSA, Ryzen AI NPU
  • Runs on macOS, iOS, Android, Linux, Windows, WebAssembly and Raspberry Pi
  • Quantized models and small footprints: tiny needs ~273 MB, large ~3.9 GB

Weaknesses

  • whisper-cli accepts only 16-bit WAV unless built with the FFmpeg option
  • Core ML, OpenVINO and ANEForge need Python-side model conversion before use
  • Each accelerator needs its own CMake flag and SDK; no single universal GPU build
  • Transcription only; no built-in diarization or translation to non-English targets is described
  • RAM ≥ 0.3 GB
  • GPU optional
  • Docker
  • Needs CMake, ffmpeg (optional, for audio conversion), CUDA / ROCm / OpenVINO / CANN / MUSA SDKs (optional)
  • Models: Whisper ggml models (tiny, base, small, medium, large-v1/v2/v3, large-v3-turbo), Q5_0 and other integer quantization, Core ML and OpenVINO encoder models