# Open-source alternatives to ElevenLabs

> Open-source alternatives to ElevenLabs that run locally, allow commercial use and need no paid API include GPT-SoVITS (speech generation and voice cloning; MIT; runs on CPU), Dia (speech generation and voice cloning; Apache-2.0; GPU, ~4.4GB+) and whisper.cpp (speech recognition; MIT; runs on CPU). None of the 22 projects we track covers all 4 ElevenLabs jobs, so most people combine two or more.

ElevenLabs: Paid voice AI platform for speech generation, voice cloning, transcription and voice agents. Pricing: Free tier (10k credits); paid plans from $6/month, or $5/month billed annually (checked 2026-10-15). Facts checked 2026-10-16.

## Projects

| Project | Jobs | Runs locally | Commercial use | Paid services | Hardware |
|---|---|---|---|---|---|
| [VoiceStudio](https://repoquarry.dev/projects/voicestudio) | speech-generation, voice-cloning, speech-recognition | Yes, desktop app and backend run on your hardware; remote services optional | No with the default OmniVoice model (non-commercial weights); app code alone allows it under AGPL | None required; VoiceStudio Pro features and a closed-source commercial licence are paid extras | NVIDIA CUDA or Apple Silicon MPS accelerated; CPU-only works but slower (~5 GB disk); Intel Mac UI only |
| [GPT-SoVITS](https://repoquarry.dev/projects/gpt-sovits) | speech-generation, voice-cloning | Yes, local WebUI on GPU, Apple silicon or CPU; Docker images available | Yes, MIT code and core weights; check licences of optional third-party models you add (UVR5, Damo ASR, Faster Whisper) | None required | No minimum RAM/VRAM stated. Stated speed (v2 ProPlus): RTF 0.028 on RTX 4060 Ti, 0.014 on RTX 4090, 0.526 on an M4 CPU. Tested on CUDA, Apple silicon and CPU. |
| [whisper.cpp](https://repoquarry.dev/projects/whisper-cpp) | speech-recognition | Yes, CPU-only or GPU, fully offline | Yes, MIT code and weights | None required | CPU fine; memory from ~273 MB (tiny) to ~3.9 GB (large) |
| [Whisper](https://repoquarry.dev/projects/whisper) | speech-recognition | Yes, CLI and Python package run on your own machine | Yes, MIT code and weights | None required | Required VRAM per model from ~1 GB (tiny, base) to ~10 GB (large); turbo ~6 GB. CPU use is not discussed in the README |
| [Fish Speech](https://repoquarry.dev/projects/fish-speech) | speech-generation, voice-cloning | Yes, CLI, API server and WebUI, or Docker; Drama 3 needs an NVIDIA GPU, S2-Pro also runs on CPU | Only with Drama 3 and only while you and your affiliates earn under USD 500k a year; S2-Pro needs a paid licence for any commercial use; offering a hosted TTS/cloning service always needs a separate licence | None for local research/non-commercial use; a paid commercial licence is needed above the revenue cap or for hosted services | Drama 3 needs at least 48 GB GPU memory (FP8), NVIDIA only; S2-Pro needs at least 24 GB, NVIDIA GPU or CPU |
| [faster-whisper](https://repoquarry.dev/projects/faster-whisper) | speech-recognition | Yes, Python library runs on CPU or NVIDIA GPU | Yes, MIT code and MIT-tagged default model weights | None required | CPU works (small model int8: ~1.5 GB RAM in README benchmark); GPU needs cuBLAS and cuDNN 9 for CUDA 12 (large-v2 int8: ~2.9 GB VRAM on an RTX 3070 Ti 8GB). These are benchmark figures, not stated minimums |
| [WhisperX](https://repoquarry.dev/projects/whisperx) | speech-recognition, speaker-identification | Yes, CLI and Python package on GPU or CPU | BSD-2-Clause code and MIT Whisper weights allow it; diarization adds CC-BY-4.0 attribution terms, and alignment model licences vary by language and are not stated | None required; diarization needs a free Hugging Face read token and accepting the pyannote model's user agreement | Under 8 GB GPU memory for large-v2 with beam\_size=5; CPU mode supported (int8); CUDA toolkit 12.8 for GPU |
| [LiveKit Agents](https://repoquarry.dev/projects/livekit-agents) | voice-agents, phone-calls | Yes; console mode runs locally, and the full stack including LiveKit server can be self-hosted; LiveKit Cloud is the hosted option | Apache-2.0 framework; bundled turn-detector models carry LiveKit Model License restrictions; plugged-in providers have their own terms | README example uses LiveKit Inference (part of LiveKit Cloud, billed by usage) or provider keys for Deepgram, OpenAI and Cartesia. LiveKit server can be self-hosted and docs cover a local Ollama LLM, but no local STT or TTS plugin is documented in the sources checked | Framework requirements not stated; the turn-detector model runs on CPU and needs under 500 MB RAM |
| [Piper](https://repoquarry.dev/projects/piper) | speech-generation | Yes, fully local on CPU, optional CUDA GPU | Possible with conditions; GPL-3.0 code and per-voice licences must be checked | None required | Runs on CPU via onnxruntime, optional CUDA; older prebuilt binaries target Raspberry Pi 3/4; no RAM numbers stated |
| [F5-TTS](https://repoquarry.dev/projects/f5-tts) | speech-generation, voice-cloning | Yes, local PyTorch on NVIDIA, AMD, Intel GPU or Apple Silicon; Docker image available | No with the released weights (CC-BY-NC-4.0); the MIT code could be used with your own commercially licensed weights | None required | Not documented; only a benchmark on a single NVIDIA L20 GPU is given |
| [Chatterbox](https://repoquarry.dev/projects/chatterbox) | speech-generation, voice-cloning | Yes, on GPU (CUDA), Apple MPS or CPU | Yes, MIT code and weights (generated audio is watermarked) | None required; Resemble AI's paid hosted TTS service is offered as an optional extra | Nano runs on CPU at about 3x realtime on 8 cores; no VRAM or RAM minimums stated |
| [Pipecat](https://repoquarry.dev/projects/pipecat) | voice-agents, phone-calls | Yes, runs on your machine; also deployable to Pipecat Cloud | Yes for the framework (BSD-2-Clause); check the licence or terms of each service you plug in | Quickstart needs Deepgram, OpenAI and Cartesia API keys; documented local alternatives include Whisper STT (Faster Whisper or MLX), Ollama LLM and Piper TTS, so a fully local pipeline is possible | Not documented for the framework; depends on the chosen services |
| [Scriberr](https://repoquarry.dev/projects/scriberr) | speech-recognition, speaker-identification, meeting-summaries | Yes, self-hosted server with a browser UI; fully offline | Yes; MIT code, MIT and CC-BY-4.0 weights with attribution | None required; OpenAI API optional for summaries and chat (free Hugging Face token for diarization) | CPU image for machines without NVIDIA GPU; CUDA images for GTX 10 to RTX 50 series. RAM not documented |
| [Bolna](https://repoquarry.dev/projects/bolna) | voice-agents, phone-calls | Yes, self-hosted with Docker; also usable as a Python package | Yes for the framework (MIT); check each provider's terms | Phone calls need a Twilio or Plivo account; ASR, LLM and TTS can be paid APIs or self-hosted endpoints | Not documented; depends on the providers you choose |
| [Orpheus TTS](https://repoquarry.dev/projects/orpheus-tts) | speech-generation, voice-cloning | Yes, GPU via vLLM, or CPU via the llama.cpp-based orpheus-cpp | Likely yes under Apache-2.0, but check the Llama 3.2 base-model terms and the model card's misuse rules | None required; Baseten hosted deployment is offered as an optional extra | GPU path via vLLM; a no-GPU path via llama.cpp is documented; no VRAM or RAM numbers stated |
| [Coqui TTS](https://repoquarry.dev/projects/coqui-tts) | speech-generation, voice-cloning | Yes, Python library and CLI on CPU or GPU, plus a CPU Docker image | Code is usable commercially under MPL-2.0, but XTTS-v2 (the main voice cloning model) is non-commercial only, including its outputs; other models carry their own licences | None required | No RAM or VRAM requirements documented; code falls back to CPU when CUDA is unavailable and a CPU Docker image exists |
| [Sesame CSM](https://repoquarry.dev/projects/sesame-csm) | speech-generation | Yes, Python code runs the model on your own machine | Apache-2.0 code and CSM weights allow it, but the required Llama 3.2 tokenizer is under the Llama 3.2 Community License (extra terms, e.g. a separate licence above 700M monthly active users), and the README describes the project as for research and educational purposes with listed prohibited uses | None required; a free Hugging Face account with access granted to CSM-1B and Llama-3.2-1B is needed | README lists a CUDA-compatible GPU as a requirement; VRAM and RAM amounts are not stated (example code also selects MPS or CPU if no CUDA) |
| [Dia](https://repoquarry.dev/projects/dia) | speech-generation, voice-cloning | Yes, on an NVIDIA GPU (CPU support not yet available) | Yes by licence (Apache-2.0 code and weights), though the README says the model is intended for research and educational use and forbids identity misuse and deceptive content | None required | NVIDIA GPU; about 4.4 GB VRAM in bfloat16/float16, about 7.9 GB in float32 (benchmarked on RTX 4090) |
| [OpenVoice](https://repoquarry.dev/projects/openvoice) | voice-cloning, speech-generation | Yes, Python package with notebooks and a local Gradio demo | Yes, MIT code and weights; V2 also needs MeloTTS, which is MIT | None required | Not documented |
| [MeloTTS](https://repoquarry.dev/projects/melotts) | speech-generation | Yes, CPU or GPU (CUDA or MPS) | Yes, MIT code and weights | None required | CPU is enough for real-time inference; no RAM or VRAM numbers stated |
| [StyleTTS 2](https://repoquarry.dev/projects/styletts2) | speech-generation, voice-cloning | Yes, Python/PyTorch inference via the provided notebooks | Code is MIT, but the pretrained weights have no formal licence (only disclosure/permission terms) and inference relies on a GPL-licensed phonemizer; an MIT-only package using gruut exists at lower quality | None required | Inference requirements not documented. CPU inference is suggested as a fix for noise on older GPUs. Fine-tuning example took about 4 hours on four NVIDIA A100s. |
| [Kokoro](https://repoquarry.dev/projects/kokoro) | speech-generation | Yes, Python library runs on your own machine | Yes, Apache-2.0 code and weights | None required | Not documented; the model is 82M parameters |

## Common questions

### Is there a free, open-source alternative to ElevenLabs?

Yes. VoiceStudio, GPT-SoVITS, Fish Speech, Chatterbox and F5-TTS are free to download and run on your own computer without paid API keys. Each covers only part of what ElevenLabs does, so check which job you need.

### Can I do voice cloning locally without an ElevenLabs subscription?

Yes, with Fish Speech (Custom licence), GPT-SoVITS (MIT) and Chatterbox (MIT). Check the hardware notes and licence conditions on each page before you start.

### Can I do speech recognition locally without an ElevenLabs subscription?

Yes, with whisper.cpp (MIT), faster-whisper (MIT) and WhisperX (BSD-2-Clause). Check the hardware notes and licence conditions on each page before you start.

### Which ElevenLabs alternatives can I use commercially?

Confirmed: Bolna, Chatterbox, Dia, faster-whisper, GPT-SoVITS, Kokoro, MeloTTS, OpenVoice, Pipecat, Scriberr, whisper.cpp and Whisper. With conditions: Coqui TTS, Fish Speech, LiveKit Agents, Orpheus TTS, Piper, Sesame CSM, StyleTTS 2 and WhisperX. Not for commercial use: F5-TTS and VoiceStudio, usually because the model weights are non-commercial. This is not legal advice; read each licence.

### Does any open-source project replace all of ElevenLabs?

Not as of 2026-10-16. Speech generation, voice cloning, speech recognition and voice agents are covered by different projects.

---
Source: RepoQuarry (https://repoquarry.dev/alternatives/elevenlabs). Data licence: CC BY 4.0. Quotes are from project documentation and belong to their authors.