Cloud vs Local Voice Cloning

What changed in 2026

For most of 2024 and 2025, the cloud-vs-local trade-off was real: cloud meant ElevenLabs-grade quality, local meant "good enough but you can hear the difference."

That changed on March 9, 2026, when Fish Audio open-sourced S2 under Apache 2.0 — the same model that ranks #1 on TTS-Arena and beat ElevenLabs V3 60/40 in their published blind A/B test. The quality gap closed. The choice is now genuinely about your situation, not a quality compromise.

Cloud vs local at a glance

Factor	Cloud	Local
Quality ceiling	Highest available	Same as Fish Audio cloud (S2)
Setup time	5 minutes	1-3 hours
Privacy	Audio leaves your machine	Stays on your hardware
Cost — low volume	Free tier or $11/mo	$1,500+ for GPU
Cost — high volume	$100-1,000+/mo	Just electricity
Maintenance	None	Updates, dependencies
Internet required	Yes	No
Vendor risk	Pricing changes, policy changes	None

When cloud still wins

• You need it working today, no setup time
• Your volume is genuinely low (< 30 minutes/month)
• You want polish features: voice library, dubbing, SFX, studio UI
• You don't have a compatible GPU and don't want to buy one
• You're testing, prototyping, or just curious

The Fish Audio free tier (8,000 credits/month with cloning) covers most of these cases for $0.

When local wins

• Privacy matters: healthcare, legal, financial, government, journalism, anything regulated
• High volume: if you're generating > 5-10 hours/month, the GPU pays back quickly
• You're building a product and don't want a vendor in your critical path
• Cost-at-scale: ElevenLabs at retail runs ~$165 per 1M chars. Self-hosted Fish runs cents
• Air-gapped or offline use cases

How to self-host Fish Speech S2

S2 is the strongest open model on the market. Setup is straightforward:

1. Hardware: RTX 4090 (or similar 24GB VRAM card). Apple Silicon works but is slower 2. Model weights: Pull from Hugging Face (`fishaudio/fish-speech-s2`) — Apache 2.0 licensed 3. Inference engine: SGLang ships in the official repo. Compose-friendly 4. Voice cloning: Provide a 15-second reference clip, generate cross-lingually in ~50 languages 5. API: SGLang exposes an OpenAI-compatible endpoint, so you can swap your ElevenLabs API URL for localhost and most apps just work

Trained on 10M+ hours, ~50 languages. The same weights powering the hosted API.

Lighter alternatives

If you don't have a 4090-class GPU:

• Qwen3-TTS — Strong quality, runs on 8GB GPUs or M-series Macs. The model behind our free tool
• F5-TTS — Fast inference, smaller footprint
• OpenVoice v2 — Good for cross-lingual cloning at lower hardware cost
• XTTS-v2 — Coqui's model, well-documented, multilingual

These are the realistic 2026 options. Older models (Tortoise, original Bark, StyleTTS2) are no longer competitive.

Our recommendation

Most individuals: Start with Fish Audio's free tier — same S2 quality, zero setup, no card. Move to a paid plan ($11/mo Plus) when you outgrow it.

Privacy-conscious users: Self-host Fish Speech S2 on a 4090 box, or Qwen3-TTS on a Mac. The Apache 2.0 license means no usage restrictions.

Builders and developers: If voice is core to your product, self-host. The cost-at-scale story compounds, and you remove a single point of failure from your stack.