The landscape for open-weight text-to-speech (TTS) has fundamentally shifted. We have reached a point where models like Kokoro-82M perform efficiently on a laptop CPU while outperforming massive proprietary models in blind listening tests. The industry has moved past the era of robotic, low-fidelity synthesis toward production-grade audio that supports zero-shot cloning, complex prosody, and fine-grained emotional control.
The Current Leaderboard
Current evaluation in the Artificial Analysis Speech Arena confirms that model size no longer dictates performance. Efficient architectures now dominate, with top-tier models ranging anywhere from 82M to 4B parameters. The economic gap remains significant: while closed-source APIs like ElevenLabs v3 cost roughly $100 per million characters, open-weight equivalents can be hosted for a marginal cost nearing zero.
| Model | Elo | License | Key Strength |
|---|---|---|---|
| Step Audio EditX | 1118 | Apache 2.0 | Emotion/Style editing |
| Fish Audio S2 Pro | 1110 | Research/Non-Comm | Inline prosody tags |
| Voxtral TTS | 1077 | CC BY-NC 4.0 | 3s Voice cloning |
| Kokoro 82M | 1060 | Apache 2.0 | CPU performance |
Notable Model Deep Dives
Step Audio EditX
Currently leading at 1118 Elo, Step Audio EditX is an audio-editing model that happens to perform TTS. It utilizes a dual-codebook tokenizer to allow for iterative style adjustments. It supports 14+ emotions and 20+ paralinguistic tags.
Fish Audio S2 Pro
This model offers perhaps the most intuitive developer experience via inline prosody tags. You can manipulate the speech output directly in the text prompt:
[clears throat] Hey, where are the docs for this API?
[whisper] I really need to get this push finished.
Voxtral TTS
Mistral’s entry is the go-to for simple 3-second voice cloning. It is built as a 4B parameter system, including a 3.4B decoder and flow-matching acoustic model, making it remarkably stable for zero-shot tasks.
Critical Licensing Caveats
Do not conflate "open weights" with "open source." Many high-performing models, including Fish Audio S2 Pro and Voxtral, carry non-commercial restrictions. If your stack requires commercial viability, prioritize:
- Step Audio EditX (Apache 2.0)
- Kokoro (Apache 2.0)
- Maya1 (Apache 2.0)
- Chatterbox (MIT)
Serving Infrastructure
Unlike the LLM world, there is no single "Ollama for TTS." Most models require specific serving paths. For production, look toward vllm-omni which provides /v1/audio/speech endpoints and WebSocket streaming. For lightweight, CPU-based inference, the kokoro-onnx ecosystem remains the most accessible path for edge projects.
If you need to expose your local TTS endpoint to external services without configuration hurdles, a simple tunnel can bridge the gap:
ssh -p 443 -R0:localhost:8880 free.pinggy.io
This command exposes your localhost port to a public HTTPS URL, which is sufficient for integrating your self-hosted model with external AI agents or client applications.
Top comments (0)