Bench
We test the pieces Holly is built from on our own Macs, and show every number and transcript.
Speech to text · October 2, 2026
Speech to text, compared
Whisper and Apple's recognizer on English, Chinese, and both in one sentence. Speed and accuracy on a MacBook Air, with every transcript.
- 5%
- errors, Whisper large v3, the most accurate on the Mac
- 1.4 s
- its typical wait per turn
- 7
- recognizers, 3 recordings, 32 turns
Text to speech · October 2, 2026
Text to speech, compared
14 voice models saying the same lines in English, Chinese, and both mixed. Speed, memory and licenses on a MacBook Air, with the clips to listen to.
- 14×
- real time for Kokoro, Holly's voice today, on the Mac's GPU
- 1.4×
- for Qwen3-TTS, which copies a voice in both languages
- 14
- models, 110 clips to hear
Voice brain · October 3, 2026
Claude models, compared
Haiku, Sonnet, Opus and Fable at three efforts, answering a spoken question the way Holly does. How soon each starts talking and whether it's right, with every reply.
- 0.73 s
- to the first word for Opus 5.5 low, Holly's pick
- 10 of 10
- right on the time sum, where Sonnet 5.5 got 7–8
- 12
- models and efforts, 350 replies to read