README.md

metadata

license: cc-by-sa-4.0
language:
  - et
tags:
  - text-to-speech
  - automatic-speech-recognition
  - piper
  - onnx
  - whisper.cpp
  - robotics

Go2 dock models

Estonian speech models that Meridein's Unitree Go2 EDU docks run: the voices (Pocket TTS, Piper) and speech recognition (Whisper). All are derived from public models, converted or packed for the dock's Jetson. A dock downloads them at a pinned commit and checks each file's sha256 before using it.

Files are never overwritten here. A new version gets a new file name, so a release that pins a commit and a sha256 keeps working.

file	what	size	sha256
pocket-cpp-et.tar.gz	Pocket TTS, Estonian, as ONNX for PocketTTS.cpp (one folder, pocket-cpp-et/)	436 064 256	1cf3ae8b3eab0724a023a8e5852272e0a870fccbd5fe7bc61a48940a680cd3ca
ggml-whisper-large-v3-turbo-et-verbatim-2604-q5_0.bin	Whisper large-v3-turbo, Estonian, q5_0 for whisper.cpp	574 041 195	ec401581b6ac1788b4eff8b92c5257a65bf24b50be1e488387fd9760166a1ad7
vits-piper-et_EE-news-medium.tar.gz	Piper's Estonian news voice (four speakers) for sherpa-onnx (one folder, vits-piper-et_EE-news-medium/)	80 299 072	55ef843f59f38a67d6acf5c6b9c62a913516784c6f529ae1c92359f61a45d927

pocket-cpp-et.tar.gz: Pocket TTS, Estonian (ONNX)

• Run by PocketTTS.cpp at e801e7d, with --precision int8. The int8 set is what runs; the fp32 files are kept beside it.
• Made from cbentes/pocket-tts-estonian: the weights and tokenizer at e8d2fd9cc01ae6e4e505269adcc1d9bdaecbea0f, the config and the six voice clips at b37e49dd65514fb79a05cff2ace696e24af15d66. Exported with pocket-tts 2.1.0 and PocketTTS.cpp's export_onnx.py at e801e7d; the beginning-of-sequence before the voice is in mimi_encoder.onnx. export_info.json inside records the same.
• Licence: CC BY-SA 4.0, as the weights it is made from, which were trained on the TalTech Estonian Speech Dataset 1.0. Credit: cbentes/pocket-tts-estonian and TalTech. Changes: converted to ONNX and quantised to int8. Share-alike: this derivative is under the same licence.

ggml-whisper-large-v3-turbo-et-verbatim-2604-q5_0.bin: Whisper, Estonian

• Run by whisper-server (whisper.cpp), language et.
• Made from TalTechNLP/whisper-large-v3-turbo-et-verbatim-2604, its ggml model (ggml/ggml-model.bin, sha256 4e231f931f029850d7ca497efe4223f85c495a27d986d955f9a577ac1cd02edb), quantised to q5_0 with whisper.cpp's quantize tool (whisper.cpp 60c0be6). The quantised model needs about 1 GB less memory.
• Licence: that of TalTechNLP's model, which this file is derived from. See their model card; credit: TalTech (Tallinn University of Technology). The repository's cc-by-sa-4.0 tag is the Pocket model's.

vits-piper-et_EE-news-medium.tar.gz: Piper, Estonian

• Run by sherpa-onnx's VITS engine; speakers Albert 0, Kalev 1, Mari 2, Vesta 3.
• Made from rhasspy/piper-voices at e7229f7d33a10e145e109b886d661984201ac3c4, et/et_EE/news/medium/: sherpa-onnx's metadata written into the .onnx, tokens.txt from the .onnx.json (the five two-letter vowels left out: unused without vowel_clusters), and the espeak-ng data from sherpa-onnx's vits-piper-en_US-lessac-medium package (it includes Estonian). Rebuilt the same, byte for byte, by tools/models/piper_et.sh in unitree-go2-dev.
• Licence: the voice's dataset is CC BY (the Speech Corpus of Estonian News Sentences, META-SHARE "Available - Unrestricted Use"; see MODEL_CARD inside). The espeak-ng data is espeak-ng's, GPL-3.0.

Checking a download

sha256sum -c SHA256SUMS