Commit History
-
web: pinyin Chinese voice bundle (zh-xiaoya, 1,554,284 params, f16)
Verified signaturec532a5d -
Chinese voice distilled from the pinyin teacher (zh-xiaoya-1p55m)
Verified signature1,554,284 parameters, 22.05 kHz, distilled from zh_CN-xiao_ya-medium. Replaces a voice at 0.468 CER that retained only 0.19x of its teacher. This one measures 0.199 on held-out FLORES and 0.024 on out-of-domain Tatoeba, retention 0.94x -- the distillation is no longer what limits Chinese. All CERs fold traditional to simplified before scoring. Without that fold every Chinese number this project recorded was inflated roughly 2x, which is why the better teacher was rejected twice before. phoneme_type is pinyin, not espeak: 85 symbols, initial/final/tone. Needs sanotts >= 0.7 for the pinyin front end.fb2394b
-
gguf: add the hi piperlite voice for audio.cpp
Verified signature7e57059 -
gguf: add the ne piperlite voice for audio.cpp
Verified signature21abfad -
gguf: nine piperlite language voices for audio.cpp (cs, de, es, fr, it, pt, ro, ru, tr)
Verified signatureFP32 GGUF plus config.json per voice, exported by tools/export_piperlite_gguf_f32.py from the roota.raw-fp16.v1 packages. Same layout as the amy/vi/id packages.6c187c0
-
Voice bundles at float16, and the eight ~511k variants
Verified signatureThe web bundles widened fp16 to f32 at export and shipped double the bytes. They now ship float16 and are widened after download; the browser verifies a sha256 of the widened result. 79.0 MB -> 39.5 MB across the eighteen. float16 is the reference here, not an approximation: every CER and WER number these voices have was measured on their roota.raw-fp16.v1 package. The twelve previously released voices are untouched and stay f32. Also adds the eight ~511k variants.94fda4d -
Add ten new-language voice bundles for the browser demo
Verified signaturede, fr, es, it, pt, ru, cs, ro, tr, ar at ~1.56M each, in the web {meta.json, front_f32.bin, dec_f32.bin} format the demo fetches. Russian and Arabic additionally need assets served same-origin by the site: web/g2p-lazy/ru (sharded espeak lexicon, ~150KB per page of prose) and web/tashkeel (the 1.16M-parameter diacritizer Piper runs before espeak; without it Arabic scores 44.55% of the correct phoneme ids).593dbac -
Add 18 voices in 10 new languages (de, tr, ru, fr, es, it, pt, ro, cs, ar)
Verified signatureTen languages distilled from Piper teachers, each shipped at ~1.56M and, for eight of them, ~511k. Measured with Whisper on held-out FLORES-200 and on out-of-domain Tatoeba text the models never saw. Out-of-domain CER: pt 0.014, it 0.026, ru 0.031, de 0.032, es 0.050, tr 0.059, ar 0.074, fr 0.088, ro 0.104, cs 0.111. The ~511k versions run 0.032-0.131 on the same set. Intelligibility only -- no listening study. These languages have no espeak-free G2P yet, so they use espeak-ng at inference.7015450
-
gguf: id config
Verified signature4c43b20 -
gguf: add id piperlite FP32 (audio.cpp package)
Verified signature5c28b4f -
gguf: vi config
Verified signatureb851a55 -
gguf: add vi piperlite FP32 (audio.cpp package)
Verified signatured163fd6 -
gguf: kristin config
Verified signature62a0d28 -
gguf: add kristin piperlite FP32 (audio.cpp package)
Verified signature7714329 -
gguf: hfc config
Verified signature2cbee3b -
gguf: add hfc piperlite FP32 (audio.cpp package)
Verified signaturee86ac6f -
gguf: amy config
Verified signature3e74916 -
gguf: add amy piperlite FP32 (audio.cpp package)
Verified signature0b2d63b -
gguf: re-embed the two-package model spec in heart-nano
Verified signature2e35dad -
gguf: heart config
Verified signature3fe79ea -
gguf: add heart 2.27M FP32 (audio.cpp package)
Verified signature639f417 -
gguf: embed the audio.cpp model spec (standalone package)
Verified signatureee8e522 -
gguf: add audiocpp exact tensor shape metadata (fixes trailing-1 dim squeeze)
Verified signaturea895eed -
FP32 GGUF with PyTorch-shaped tensors, for ggml consumers
Verified signaturea41ee1b -
FP32 GGUF with PyTorch-shaped tensors, for ggml consumers
Verified signature8c114e8 -
card: npm 0.3.0 streams weights from here too
Verified signaturef11e871 -
web: browser-format voice blobs so sanotts-web can fetch weights from here
Verified signaturebc04c49 -
config.json: carry tensor->offset regions so consumers need no name translation
Verified signature8724338 -
Add sanoTTS GGUF config.json for audio.cpp
Verified signature939137c -
Add sanoTTS nano GGUF packaging for audio.cpp
Verified signature41fb858 -
Add sanoTTS nano GGUF packaging for audio.cpp
Verified signatureff00c41 -
Add sanoTTS nano GGUF packaging for audio.cpp
Verified signaturedbcbdc3 -
samples: drop the mislabelled 745k filename, superseded by mcu-567k.mp3
Verified signature753380a -
card: 294k-2.27M range, heart + heart-nano, HF-first downloads, 567,008 on-device count
Verified signature8ba4acb -
evidence: diverse24 re-measurement for heart and heart-nano
Verified signatured28fbbb -
samples: rename mcu-745k to the audited 567,008 count
Verified signaturee4f82b2 -
samples: add heart (2.27M) and heart-nano (294k) clips
Verified signature11aecf1 -
Add heart voice package (nano runtime: meta.json + offsets header + int8/fp32 blobs)
Verified signaturea6126c1 -
Add heartnano voice package (nano runtime: meta.json + offsets header + int8/fp32 blobs)
Verified signature4b11018
-
Rebuild model card: samples table with embedded audio players, install table, links
Verified signaturedd52a31 -
Add voice sample audio clips (16 language/voice + mcu-745k)
Verified signature19c75fd -
sanoTTS voice packages: 7 voices, 6 languages, 1.08M-1.84M params
Verified signaturec79c043 -
initial commit
Verified signaturecbf85e0