speakilo-cpu-models / README.md
aaronpham's picture
Document ASR and MT assets
a00c473 verified
|
Raw History Blame Contribute Delete
2.39 kB
---
library_name: onnxruntime
tags:
- speech
- cpu
- vad
- language-identification
- text-to-speech
---
# Speakilo CPU speech assets
Private deployment bundle for Speakilo's CPU VAD, LID, and TTS stages.
This repository mirrors third-party artifacts; it does not claim authorship
or grant rights beyond each upstream project's terms.
## Layout
- `vad/silero/`: Silero VAD v6 ONNX asset used by the installed runtime.
- `lid/voxlingua107-ecapa/`: SpeechBrain VoxLingua107 ECAPA checkpoint and configuration.
- `tts/`: five Speakilo CPU voice bundles and their required frontend/runtime files.
- `checksums.json`: SHA-256, byte size, source path, and upstream metadata.
## Upstream sources and terms
- Silero VAD: `snakers4/silero-vad` (MIT upstream project).
- VoxLingua107 ECAPA: `speechbrain/lang-id-voxlingua107-ecapa`
at `0253049ae131d6a4be1c4f0d8b0ff483a0f8c8e9` (Apache-2.0 metadata).
- Ppaso Korean: `akamotaco/ppaso-tts-v1`
at `53d09664c4f636a5fb6f2ebe3ec22cd83ee249b9` (Apache-2.0; bundled notices apply).
- Piper voices: `rhasspy/piper-voices`; preserve each included model card.
Dataset terms differ: VI 25-hours is marked unknown, Lessac has research-use
lineage, and VAIS1000 is CC-BY-4.0 but fine-tuned from Lessac.
- AISHELL3 bundle: from the sherpa-onnx TTS release. The downloaded bundle has
no verified model/data license. Keep this repository private unless rights
are independently established.
## Runtime
The Piper/AISHELL models use sherpa-onnx VITS on CPU. Ppaso uses its ONNX CPU
adapter. VoxLingua LID uses SpeechBrain/PyTorch CPU. Silero runs as ONNX.
The files are mirrored byte-for-byte from the local Speakilo installation.
See `checksums.json` before deployment.
## ASR and MT additions
- `asr/gipformer/`: Vietnamese GIPFormer 1.5 68M RNN-T INT8,
`g-group-ai-lab/gipformer1.5-68M-rnnt@dd9227dcd8705c13f33bdbe59728d546ab94480f`
(MIT metadata).
- `asr/sensevoice-sherpa-int8/`: SenseVoiceSmall sherpa-onnx INT8,
`csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17@2365baeacb507f821a0c8120fcee3d484dba7a07`.
Upstream Hugging Face metadata does not declare a license; keep private.
- `mt/hy-mt2/Hy-MT2-1.8B-Q4_0-imat-q8emb-allq40.gguf`: user-supplied
GGUF derived from `tencent/Hy-MT2-1.8B@9a341cd1b679d3efd23b46e847b01745a71ed792`
(Apache-2.0 base-model metadata), intended for the llama.cpp Hexagon/HTP runtime.