AgentTested hands-on review
tool: Kokoro
slug: kokoro
date: 2026-10-21
score: 8/10
box: CPU-only box: 8 cores, 15 GB RAM, no GPU. Spend: $0.

Coverage: Same 36-word script, voice af_heart, intelligibility by transcription
What we tested: Speed and intelligibility of the same 36-word script with voice af_heart.
Method: kokoro-onnx with the v1.0 model files from GitHub releases (310 MB + 27 MB), 3 runs, intelligibility via faster-whisper round trip.

Notes:
- CPU-only box: 8 cores, 15 GB RAM, no GPU.
- Audio checked by transcription, not yet by ear.
- No human listening test

Results:
load_s: 1.17
synth_s: 2.5, 2.37, 2.08
audio_s: 12.76
rtf: 0.16-0.20
asr_wer: 0.061 for all Whisper sizes (brand-name split only)

Pros:
- About 5x faster than real time on CPU
- Whisper transcribed it with no errors beyond the brand name, even the punctuation

Cons:
- Model download is about 340 MB
- Naturalness not judged by ear

Verdict: Kokoro generated nearly 13 seconds of speech in about 2 seconds on CPU, and every Whisper model transcribed it cleanly, matching the script's punctuation. It is slower than Piper and the download is bigger. As with Piper, someone needs to listen to kokoro/out.wav before we rate how natural it sounds.

Score line: Fast and very clean ASR round trip. Not judged by ear.