Run Without the Fish Sidecar

For single-GPU hosts, or when you don't need voice cloning, skip the Fish Audio container entirely.

One-service start

TTS_BACKEND=kokoro docker compose up -d protovoice

• protovoice alone. No fish-speech.
• Kokoro runs in-process inside the protovoice container (~2 GB VRAM, no GPU sharing penalty).
• FISH_URL is ignored.

VRAM budget (single GPU)

Component	VRAM
Whisper large-v3-turbo	~6 GB
Routing vLLM (Qwen 4B)	~15 GB
Kokoro 82M	~2 GB
Total	~23 GB

Fits on a 24 GB+ card (RTX 3090, 4090, PRO 6000, A100, H100, etc.).

Picking a Kokoro voice

TTS_BACKEND=kokoro KOKORO_VOICE=am_michael docker compose up -d protovoice

Common voices:

• af_heart (default, American female)
• af_bella, af_nicole, af_sarah
• am_adam, am_michael
• bf_emma, bf_isabella (British female)
• bm_george, bm_lewis (British male)

Full list on the Kokoro HF card.

Trade-offs

	Fish S2-Pro	Kokoro 82M
First-token latency	~400-800 ms steady-state	~50 ms
Voice cloning	✓ 10-30 s reference	✗ fixed voices
Prosody control tags	✓ 15 k+ tags	✗
Naturalness	Excellent	Good
Cold-start compile	~2 min	~2 s
GPU footprint	+22 GB on a separate GPU	+2 GB in-process
Extra container	Yes	No

If you care about turn latency and don't need cloning, Kokoro is often the better choice.