Chatterbox Finnish Nano v0.1.3
Finnish-only adaptation of Chatterbox Nano for conversational TTS.
The default built-in Nano voice works without a reference recording, and optional
reference conditioning is also supported.
Release: v0.1.3
Canonical T3: FP16
T3 SHA256: d25e3ac95eefefb58c40ad4d5a47b4fa621bc1bd92e91f2d904f622d5458ad26
Automatic evaluation
v0.1.3 final release gate was evaluated at FP16 runtime precision.
Finnish quality set:
| Metric | Result |
|---|
| Prompts | 12 |
| WER | 0.0822 |
| CER | 0.00949 |
| Exact transcript | 9/12 (75.0%) |
Finnish EOS-ASR stress set:
| Metric | Result |
|---|
| Generations | 100 |
| WER | 0.0848 |
| CER | 0.01382 |
| Exact transcript | 60/100 (60.0%) |
Endpoint safety gate used 100 Finnish generations plus 100 English
regression-canary generations:
- generation-limit hits: 0
- premature >1 s: 0
- audio tail >1 s: 0
- audio tail >2 s: 0
- maximum measured tail: 0.78 s
English was used here only as a regression canary; this remains a Finnish-only release.
Finnish text normalization
The companion runtime enables Finnish number-to-speech normalization by default.
This is a Finnish-only release and has no language selector.
Runtime package
The pinned Nano runtime uses s3gen_meanflow.safetensors. The unused legacy
s3gen.safetensors is intentionally not shipped in v0.1.2. All other files from
the otherwise identical full v0.1.2 package are retained, including
t3_nano_v1.yaml.
Links
Provenance
Base revision: 71ccd1d0081b430592cea481f4307e764e07bc64
Chatterbox runtime revision: 5de7a54aa4e5e2baadb0182dde554908b48b85c2
Clean target source SHA256: 56b72efe90369e50e0eea35af2ea2fca540a74d84fdbfbd8bc7ef57b0fd4841d
