Qwen3-TTS vs Kokoro: which self-hosted OpenAI-compatible TTS to pick

Description and use

Qwen3‑TTS offers higher voice quality and expressiveness but needs more memory and time, whereas Kokoro‑82M is faster and lighter, suitable for real‑time use.

Problem

Choosing a self-hosted TTS model that balances voice quality, expressiveness, latency, and memory footprint.

Outcome

Qwen3-TTS delivers superior voice quality and expressive controls at the cost of higher latency and memory usage, while Kokoro-82M offers faster, low-footprint inference suitable for real-time or resource-constrained scenarios.

Workflow

• 1. Deploy Kokoro-82M or Qwen3-TTS via Railway one-click templates or Docker.
• 2. Measure latency and memory usage on target hardware.
• 3. Compare voice quality and expressiveness (e.g., emotion markup, multi-speaker presets).
• 4. Choose Qwen3-TTS for higher quality and expressiveness, or Kokoro for real-time speed and low memory.
• 5. Switch between models by changing the base URL of the OpenAI-compatible /v1/audio/speech endpoint.

Evidence

Qwen3-TTS wins, clearly. I ran both directly against each other. It's a 1.7B model versus Kokoro's 82M, and it shows — tone and pacing hold together much better across longer narration.

Source

Kokoro is faster than realtime on a shared vCPU and needs about 1GB RAM. Qwen3-TTS is CPU-bound — realtime factor around 2, so not real-time — and wants 4-6GB RAM.

Source