Kokoro vs Piper
Both are small, CPU-friendly TTS engines. On our 8 vCPU x86 test box Piper (medium voice) was about 7x faster than Kokoro fp32; Kokoro is the newer, larger neural model. The license is the big practical difference.
Side by side
| Kokoro 82M | Piper (voices) | |
|---|---|---|
| Organization | hexgrad | Rhasspy / Open Home Foundation |
| Parameters | 82M | Unknown |
| License | apache-2.0 source | Voices repo: MIT; engine (piper1-gpl): GPL-3.0 source |
| Commercial use | Yes Apache-2.0; card welcomes production/commercial deployment | Conditional Engine is GPL-3.0; each voice's training data has its own license, check the voice's MODEL_CARD |
| Stated memory | Not stated | Not stated |
| Runs on CPU | Yes | Yes |
| Streaming | Unknown | Unknown |
| Voice cloning | Unknown | Unknown |
| Languages | 9 languages | 35 languages |
| Our fastest CPU result | RTF 0.1234 kokoro-onnx fp32 (onnxruntime CPU) voice=af_heart; Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU | RTF 0.0174 piper-tts (onnxruntime) voice=en_US-lessac-medium; Cloud x86 VM, 8 vCPU Intel Xeon, 16 GB RAM, no GPU |
| Pi 5 | CPU supported, no device data | Sourced statement |
| CPU only | Measured by us | Measured by us |
| Apple Silicon | Sourced statement | CPU supported, no device data |
| 8 GB GPU | Estimate from parameter count | Unknown |
| 12–16 GB GPU | Estimate from parameter count | Unknown |
| 24 GB+ GPU | Estimate from parameter count | Unknown |
Each fact is sourced on the model's page. "Unknown" means the source doesn't say. Accuracy: Hugging Face Open TTS Leaderboard.
Key differences (sourced)
- Kokoro is Apache-2.0, and the model card welcomes commercial deployment. source
- Piper's voices repo is MIT but the current engine (piper1-gpl) is GPL-3.0, so linking it into closed-source software has copyleft implications. Check each voice's model card too. source
- Piper is the default local TTS in Home Assistant and is 'optimized for the Raspberry Pi 4'. source
- Kokoro has an ONNX build (kokoro-onnx) with fp32 and int8 files. On our box int8 was not faster than fp32. source
Which should you pick?
- Pick Piper for the lowest latency on weak hardware (Pi, N100) and for many languages and voices (35 languages per our data).
- Pick Kokoro if you need a permissive license for a commercial product and can afford roughly 7x more CPU per second of audio than Piper.
Sponsor
Reach developers picking a speech model to self-host. Advertise here →
Sponsors never affect rankings, picker results or benchmark numbers.