ElevenLabs gives you 10,000 free credits/month (~10,000 characters), then charges $6/mo for 30,000 credits (Starter), $22/mo for 121,000 credits (Creator), and $99/mo for 600,000 credits (Pro), prices checked in October 2026 on the official page. Three limits pushed us to test alternatives: your data stays on ElevenLabs servers, pro voice cloning requires the Creator plan (121 k credits), and heavy usage makes the bill climb fast (e.g., 1 M characters ≈ $18/mo beyond the Pro plan).

We ran a dozen open-source models on GPU (RTX 4090, 24 GB) and CPU (Apple M3, 16 GB) for two weeks. Here is what we kept in October 2026, with prices, licenses, and limits verified on official repos and websites.

Summary table#

Notre choixTortoiseBarkCoqui TTSPiperRVCGPT-SoVITSFish SpeechCosyVoiceOpenVoiceChatTTSMeloTTSF5-TTSEmotiVoice
Best forMax quality, multi-voiceGenerative audio (speech, music, SFX)Complete toolkit (1100+ languages)Fast local TTS, CPUReal-time voice conversionFew-shot cloning (1 min)SOTA quality, multi-characterStreaming, 9 langs + dialectsInstant cloning, multi-lingualDialogue, fine prosodyMultilingual CPU TTSFlow matching, speedEmotion, 2000+ voices
Code licenseApache-2.0MITMPL-2.0MITMITMITFISH AUDIO COMMUNITYApache-2.0MITAGPL-3.0MITMITApache-2.0
Weights licenseApache-2.0MITMPL-2.0 / CoquiMITMITMITFISH AUDIO COMMUNITY / CC-BY-NCApache-2.0MITCC BY-NC 4.0MITCC-BY-NCApache-2.0
Commercial use✅✅✅ (check XTTS)✅✅✅⚠️ See license✅✅❌ Non-commercial✅⚠️ Non-commercial✅
Voice cloning✅ (reference)❌ (presets only)✅ (XTTS v2, YourTTS)❌✅ (conversion)✅ (5 s zero-shot, 1 min few-shot)✅ (10-30 s)✅ (zero-shot)✅ (zero-shot)⚠️ Random speakers❌✅ (reference)✅ (personal cloning)
LanguagesMostly EN13+1100+ (Fairseq)100+Model-dependentEN, JA, KO, YUE, ZH100+ (S2), 24 (Drama 3)9 + 18 CN dialectsEN, ES, FR, ZH, JA, KOEN, ZHEN, ES, FR, ZH, JA, KOEN, ZHEN, ZH
Min VRAM (GPU)4 GB (streaming)2-12 GB4-8 GBCPU only4-6 GB6-8 GB8-12 GB (S2), 24 GB+ (Drama 3)4-6 GB4-6 GB4 GBReal-time CPU8 GB (L20)6-8 GB
Speed (RTF)0.25-0.3 (4 GB)~1x (GPU)Variable< 0.1x (CPU)Real-time0.014 (4090), 0.5 (M4 CPU)0.195 (H200, S2)0.15 (streaming)< 1x~0.3 (4090)Real-time CPU0.04 (L20 TRT)~1x
InterfaceCLI, API, HF SpacePython, HF Space, ColabPython, CLI, Docker, ServerCLI, Python, ONNXWebUI (Gradio)WebUI (Gradio)WebUI, API, CLI, DockerWebUI, API, vLLM, TRT-LLMPython, Colab, DockerWebUI, CLI, APIPython, WebUI, CLIGradio, CLI, DockerStreamlit, OpenAI API
GitHub stars~15k~39k~46k~11k~39k~64k~33k~24k~38k~40k~8k~16k~9k
GitHub stars rounded, licenses, and minimum VRAM checked in October 2026

Our picks in detail#

1

OpenVoice V2

Our pickOfficial site of OpenVoice V2
License
MIT (code + weights)
Hosting
Local, Docker, Colab
Price
Free, commercial use allowed

The easiest way to start without a dedicated GPU. OpenVoice V2 (April 2024) clones a voice in seconds from 5 seconds of reference audio, no training required. It natively supports English, Spanish, French, Chinese, Japanese, and Korean, with control over accent, emotion, rhythm, and pauses. The MIT license allows unrestricted commercial use (LICENSE). On CPU (Apple M3), inference takes ~3-5 s for a short sentence; on RTX 4090 it's near-instant. No official WebUI, but community Colab demos (official Colab) and an OpenAI-compatible API wrapper exist. Limitation: no few-shot fine-tuning to improve similarity, and quality sits slightly below custom-trained models.

2

GPT-SoVITS v5

Official site of GPT-SoVITS v5
License
MIT (code + weights)
Hosting
Local (WebUI), Docker, AutoDL
Price
Free, commercial use allowed

Best quality/effort ratio for personal voice cloning. With 1 minute of audio, the SoVITS fine-tuning produces a voice nearly indistinguishable from the original. Zero-shot (5 s) is already impressive. The WebUI bundles audio slicing, ASR (FunASR for Chinese, Faster-Whisper for others), labeling, vocal/accompaniment separation (UVR5), and training in a few clicks. v5 (2025) improves similarity without fine-tuning, reduces high-frequency artifacts, and adds cuda_graph + flash_attention acceleration. On RTX 4090: RTF 0.014; on Mac M4 CPU: RTF 0.5. Supports EN, JA, KO, YUE, ZH. Unambiguous MIT license (LICENSE). Quick install:

bash
conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
bash install.sh --device CU128 --source HF --download-uvr5
python webui.py
3

Fish Speech S2-Pro / Drama 3

Official site of Fish Speech S2-Pro / Drama 3
License
Code Apache-like, weights FISH AUDIO COMMUNITY / CC-BY-NC
Hosting
Local, Docker, vLLM, SGLang, cloud API
Price
Free locally, paid API on fish.audio

SOTA open-source quality, but restrictive license. S2-Pro (4B params) beats closed models on Seed-TTS Eval (WER 0.54% zh, 0.99% en) and Audio Turing Test (0.515) per the technical report. Drama 3 (35B-A3B, 3B active) adds local editing (replace a word), multi-turn revision ("gentler, sadder"), multi-character streaming production, and 24-language support (Drama 3 report). Weights on Hugging Face for both. License: local use free, commercial redistribution of weights prohibited (LICENSE). Cloud API on fish.audio (paid). vLLM/SGLang deployment recommended for production:

bash
pip install sglang[all]
python -m sglang.launch_server --model-path fishaudio/fish-speech-1.4 --port 8000
4

CosyVoice 3.0 (Fun-CosyVoice3-0.5B)

Official site of CosyVoice 3.0 (Fun-CosyVoice3-0.5B)
License
Apache-2.0 (code + weights)
Hosting
Local, Docker, vLLM, Triton/TensorRT-LLM
Price
Free, commercial use allowed

The production-ready streaming choice. Fun-CosyVoice 3.0 (Dec 2025) covers 9 languages + 18 Chinese dialects, with zero-shot cloning, pronunciation inpainting (pinyin/CMU), text normalization without a classic frontend, and bi-streaming (text-in / audio-out) at 150 ms latency per the README. 0.5B models are easy to deploy; vLLM 0.11+ and TensorRT-LLM (4x faster) supported. Apache-2.0 without ambiguity (LICENSE). Ideal for real-time voice assistants. Install:

bash
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
cd CosyVoice
conda create -n cosyvoice python=3.10
conda activate cosyvoice
pip install -r requirements.txt
python vllm_example.py  # for vLLM
5

Piper (OHF-Voice/piper1-gpl)

Official site of Piper (OHF-Voice/piper1-gpl)
License
GPL-3.0
Hosting
Local, CLI, Python, ONNX, Docker
Price
Free, commercial use allowed

Fastest TTS on CPU, no GPU required. The original rhasspy/piper (11.3k ⭐) was archived in October 2025; development continues under OHF-Voice/piper1-gpl (GPL-3.0). Hundreds of pre-trained voices (ONNX) cover 100+ languages. Inference < 0.1x real-time on modern CPU (Apple M3, Raspberry Pi 4). No voice cloning, no fine-grained emotion: pure, reliable TTS for reading articles, docs, notifications. Easy integration via piper-tts (PyPI) or standalone binaries:

bash
pip install piper-tts
echo "Hello world" | piper --model en_US-lessac-medium.onnx --output_file output.wav
6

MeloTTS

Official site of MeloTTS
License
MIT
Hosting
Local, Python, WebUI, CLI, Docker
Price
Free, commercial use allowed

High-quality multilingual TTS, real-time on CPU. Developed by MIT + MyShell. 6 languages (EN US/UK/IN/AU/Default, ES, FR, ZH mixed EN, JA, KO) with multiple English accents. The Chinese speaker handles EN/ZH code-switching. Real-time CPU inference (RTF < 1). No cloning, no parametric emotion, but excellent synthesis quality for narration, accessibility, assistants. Pure MIT (LICENSE).

bash
pip install git+https://github.com/myshell-ai/MeloTTS.git
python -m melo.api --text "Hello world" --language EN --speaker EN-US --output output.wav
7

Coqui TTS (XTTS v2, YourTTS, VITS, Tortoise, Bark…)

Official site of Coqui TTS (XTTS v2, YourTTS, VITS, Tortoise, Bark…)
License
MPL-2.0
Hosting
Local, Python, CLI, Docker, Server
Price
Free locally, Coqui cloud plans paid

The most complete toolkit (46k ⭐). 1100+ languages via Fairseq/VITS, XTTS v2 (13 languages, multilingual cloning), YourTTS (EN/FR/PT), Tortoise and Bark integrated. Unified Python API (TTS("tts_models/multilingual/multi-dataset/xtts_v2")), tts CLI, HTTP server, Docker. MPL-2.0: code is free, but Coqui now monetizes hosting and support. XTTS v2 needs ~6 GB VRAM; YourTTS ~4 GB. Note: Coqui has pivoted to SaaS, open-source maintenance has slowed.

bash
pip install coqui-tts
tts --text "Hello world" --model_name "tts_models/multilingual/multi-dataset/xtts_v2" --speaker_wav reference.wav --out_path output.wav
8

Bark (Suno)

Official site of Bark (Suno)
License
MIT
Hosting
Python, HF Space, Colab, Transformers
Price
Free, commercial use allowed

All-in-one generative audio: speech, music, sound effects, emotions. Not a classic TTS: GPT-style model generating raw audio (EnCodec) from text. 13+ languages, 100+ speaker presets. Supports tags [laughs], [sighs], ♪ lyrics ♪, [MAN]/[WOMAN]. Limit ~13-14 s per generation (context window). 12 GB VRAM for full model, "small" variants for 8 GB. No custom cloning (presets only). MIT, commercial use allowed.

bash
pip install git+https://github.com/suno-ai/bark.git
python -c "from bark import generate_audio, SAMPLE_RATE; import scipy; scipy.io.wavfile.write('out.wav', SAMPLE_RATE, generate_audio('Hello [laughs] world'))"
9

F5-TTS / E2 TTS

Official site of F5-TTS / E2 TTS
License
MIT (code), CC-BY-NC (weights)
Hosting
Local, Gradio, CLI, Docker, Triton/TensorRT-LLM
Price
Free locally, commercial use of weights not allowed

Flow matching + DiT, very fast in optimized deployment. F5-TTS v1 (Mar 2025): Diffusion Transformer with ConvNeXt V2. RTF 0.0394 on L20 (TensorRT-LLM), 0.1467 PyTorch offline. E2 TTS: Flat-UNet variant. Sway Sampling improves quality at inference. Gradio WebUI (multi-style, multi-speaker, voice chat with Qwen2.5). Models on HF (Emilia-ZH-EN). Weights license CC-BY-NC → no commercial use of pre-trained models (LICENSE).

bash
pip install f5-tts
f5-tts_infer-gradio  # launches web UI
# or CLI:
f5-tts_infer-cli --model F5TTS_v1_Base --ref_audio ref.wav --ref_text "reference text" --gen_text "text to generate"
10

EmotiVoice (NetEase Youdao)

Official site of EmotiVoice (NetEase Youdao)
License
Apache-2.0
Hosting
Docker, Streamlit, OpenAI-compatible API
Price
Free, commercial use allowed

2000+ pre-trained voices + prompt-controlled emotion. English and Chinese. Emotions: happy, excited, sad, angry, etc. OpenAI-compatible TTS API (since Dec 2023), 13,000 free calls on their endpoint. Personal voice cloning via DataBaker/LJSpeech recipes. Docker GPU required (no official CPU version). Development slowed (last commit 2024), but solid base for emotional prototypes.

bash
docker run -dp 127.0.0.1:8501:8501 -p 127.0.0.1:8000:8000 syq163/emoti-voice:latest
# OpenAI-compatible API at http://localhost:8000
11

ChatTTS

Official site of ChatTTS
License
Code AGPL-3.0, Weights CC BY-NC 4.0
Hosting
WebUI, CLI, API, vLLM (experimental)
Price
Free, **commercial use forbidden**

Best for conversational prosody (laughter, pauses, interjections). Trained on 100k+ hours (40k h open source). Fine control: [oral_2][laugh_0][break_6], [uv_break], [laugh], [lbreak]. Excellent WER on Seed-TTS. But weights license CC BY-NC 4.0 = prohibited in commercial production. AGPL-3.0 on code. Reserve for research, internal prototypes, non-monetized personal projects.

12

RVC (Retrieval-based Voice Conversion)

Official site of RVC (Retrieval-based Voice Conversion)
License
MIT
Hosting
Local WebUI, Docker, real-time (VST)
Price
Free, commercial use allowed

Real-time voice conversion (voice changer), not direct TTS. Trains a model in ≤10 min with 10 min of clean audio. Inference 170 ms (90 ms with ASIO). Full WebUI (training, inference, model merging, UVR5 separation). Used for dubbing, streaming, accessibility. Requires an upstream TTS (e.g., Piper, MeloTTS) to generate source speech. Pure MIT.

How to choose#

  • You want to clone your voice right now, no GPU, commercial use → OpenVoice V2. Free Colab, 5 s reference, 6 native languages, MIT.
  • You have a GPU (6-8 GB) and want the best personal cloning → GPT-SoVITS v5. 1 min audio → auto fine-tuning, all-in-one WebUI, MIT.
  • You target SOTA quality for published content (podcast, video) → Fish Speech S2-Pro (local) or Drama 3 (if 24 GB+ VRAM). Check license for redistribution.
  • You're building a real-time voice assistant (streaming, low latency) → CosyVoice 3.0. Bi-streaming 150 ms, instruct (emotion, speed), Apache-2.0, vLLM/TRT-LLM ready.
  • No GPU, just reliable multilingual narration TTS → Piper (CPU, 100+ langs) or MeloTTS (CPU real-time, 6 langs + accents).
  • You need marked emotions (laughter, anger, sadness) in EN/ZH → EmotiVoice (2000+ voices, OpenAI-compatible API, Apache-2.0).
  • You research conversational prosody → ChatTTS (best dialogue prosody), but non-commercial.
  • You want real-time voice changing (streaming, gaming, accessibility) → RVC + a source TTS (Piper/MeloTTS).

What benchmarks don't tell you#

Advertised VRAM ≠ real VRAM. READMEs often give the minimum to load weights; in practice, add 2-4 GB for context, KV cache, batch size. On 8 GB, prefer "small" / quantized models (Bark small, CosyVoice 300M, GPT-SoVITS half-precision).

Speed depends on length. RTF (Real-Time Factor) measured on short sentences. On long-form (book chapter), chunking, KV cache re-encoding, and memory management change the game. Fish Speech S2-Pro and CosyVoice 3.0 handle long streaming natively; Tortoise, Bark, ChatTTS need manual chunking.

Cloning quality depends on reference audio. 10 min clean (no noise, single speaker, 22-48 kHz) > 30 min noisy. For GPT-SoVITS and Fish Speech, 1-2 min suffices if clean. For OpenVoice, 5-10 s works but timbre similarity remains imperfect.

Weight licenses ≠ code licenses. This is trap #1 in 2026. Check each model: LICENSE (code) vs MODEL_LICENSE / LICENSE in weights folder / HF card. "Open source" on the README doesn't guarantee commercial freedom for weights.

Our pick by profile#

Our pick

OpenVoice V2

8.5/10

OpenVoice V2 wins on accessibility: zero GPU setup, 5 s cloning, 6 languages, MIT. GPT-SoVITS v5 is our "local quality" pick if you have 6-8 GB VRAM and 1 min of clean audio. Fish Speech S2-Pro for SOTA quality (check license). CosyVoice 3.0 for production streaming. Piper / MeloTTS for pure CPU TTS. EmotiVoice for EN/ZH emotion. ChatTTS for dialogue prosody (research only).

Who it's for
Developers who want instant voice cloning without a GPU, with a clear MIT license for commercial use.
Who should skip it
Those who need maximum timbre similarity: GPT-SoVITS v5 or Fish Speech S2-Pro after fine-tuning.

Frequently asked questions#

Frequently asked questions

What is the best free alternative to ElevenLabs for voice cloning?

OpenVoice V2 for simplicity (5 s, Colab, MIT). GPT-SoVITS v5 for quality (1 min, auto fine-tuning, WebUI, MIT) if you have a 6-8 GB GPU.

Can these models be used in commercial production?

Yes for OpenVoice V2, GPT-SoVITS, CosyVoice, Piper, MeloTTS, EmotiVoice, RVC, Tortoise, Bark (MIT/Apache-2.0 on code and weights). No for ChatTTS (weights CC BY-NC 4.0) and F5-TTS (weights CC-BY-NC). Caution for Fish Speech (restrictive community license) and Coqui XTTS v2 (check current Coqui terms).

Which model for a real-time voice assistant (low latency)?

CosyVoice 3.0: bi-streaming 150 ms, vLLM/TensorRT-LLM support, native instruct (emotion, speed). Alternative: Fish Speech S2-Pro deployed via SGLang/vLLM (RTF 0.195 on H200).

Is ElevenLabs still better than open source?

On convenience (single API, no infra, pro voices ready, dubbing studio): yes. On raw cloning quality: Fish Speech S2-Pro/Drama 3 and fine-tuned GPT-SoVITS v5 match or beat it. On cost at volume: open source local wins from ~500k chars/mo (amortized GPU cost).

How much VRAM do I really need for GPT-SoVITS v5?

6 GB (half-precision, is_half=true) for zero-shot inference. 8 GB recommended for SoVITS fine-tuning (1 min). 4 GB possible with CPU offload but slow. The WebUI handles auto-offload.

Can I do multilingual TTS (FR, EN, ES, DE, JA, KO…) open source?

Yes. CosyVoice 3.0: 9 languages + dialects. Fish Speech S2-Pro: ~100 languages (MiniMax benchmark 24 langs). Coqui XTTS v2: 13 languages. MeloTTS: 6 languages + accents. OpenVoice V2: 6 native languages. EmotiVoice: EN/ZH only.

Sources#

Official pages and repos checked on October 26, 2026.