Finnish sanoTTS RC1 runtime contract
Format: roota.raw-fp16.v1
Input boundary: Piper phoneme IDs matching piper-phoneme-config.json.
Sample rate: 22050 Hz
Components:
- duration: 135556 parameters
- acoustic: 3490201 parameters
- decoder: 730721 parameters
- total: 4356478 parameters
Acoustic config architecture: calibrated
The calibrated acoustic wrapper uses the checkpoint config and its
depthwise_lowrank output adapter. The latent boundary remains 192 channels.
The tensor payload is a single contiguous little-endian fp16 blob. Every tensor
has shape, byte offset, byte length, and SHA-256 in manifest.json.