French speech without a GPU requirement.
What ran
All three prompts generated valid audio: a welcome, payment acknowledgement and team introduction. Generation ranged from 1.98 to 3.41 seconds for 3.98–6.38 seconds of audio.
Scope of the decision
Kokoro became the French TTS pilot option. This initial smoke test does not establish African French accent quality. The later Azure CPU study repeats matched French prompts and provides a stronger hardware comparison with listening samples.
Results, in context.
| Prompt | Generation (s) | Audio (s) |
|---|---|---|
| Welcome | 2.35 | 4.62 |
| Payment | 1.98 | 3.98 |
| Introduction | 3.41 | 6.38 |
One sample per prompt; load time is excluded.
How we ran it.
- Kokoro on CPU using the ff_siwis French voice. Three synthetic service messages; one generation per prompt. Model load was measured separately at 1.36 seconds.
The limits matter.
- Three synthetic prompts, no blind listening assessment and no accent-specific validation. A successful audio file is not proof of complete or natural pronunciation.
Evidence & sources.
The public extract contains the methods, selected results and caveats in this note. It is a curated summary, not the full raw run. Fingerprints identify the retained source artifacts; gated audio and private operational records are not included.
Download results extractSource artifact fingerprints · SHA-256
72c1402af75adc57370aeb3d40fccedbfb6daed7c38e621a01a42b041df60a71Suggested citation: Embiro Labs Research. “Bonjour, Kokoro.” Published 12 Oct 2026. embiro.ai/research/kokoro-french-voice
Report a correction
