Models and voices: licenses
Homer Node downloads speech recognition and synthesis models itself, according to the node's roles. This page lists where each model comes from and under what terms it is used. Models are not part of the installer: the node downloads them from the sources below and checks their checksums.
Speech recognition
| Model | Language | License | Source |
|---|---|---|---|
| GigaAM v3 (sherpa-onnx) | Russian | MIT | salute-developers/GigaAM, packaged by sherpa-onnx |
| Parakeet TDT CTC 110M (sherpa-onnx, int8) | English | CC BY 4.0 | nvidia/parakeet-tdt_ctc-110m, packaged by sherpa-onnx |
| Vosk small ru 0.22 | Russian | Apache 2.0 | alphacephei.com/vosk |
| Vosk small en-us 0.15 | English | Apache 2.0 | alphacephei.com/vosk |
Attribution: the English speech recognition model is “Parakeet TDT CTC 110M” © NVIDIA, licensed under CC BY 4.0; Homer uses it as converted to ONNX and quantized to int8 by the sherpa-onnx project.
Voices
| Voice | Language | License | Source |
|---|---|---|---|
| Piper kristin, norman, john (medium) | English | MIT; LibriVox recordings are public domain | rhasspy/piper-voices |
| Your own voice: Qwen3-TTS 1.7B (fine-tuned for Russian stress) | Russian | Apache 2.0 | siriusfreak/qwen3-tts-12hz-1.7b-ru-stress-cf |
| Your own voice: Qwen3-TTS 1.7B Base | English | Apache 2.0 | Qwen/Qwen3-TTS-12Hz-1.7B-Base |
| Built-in “Homer” voice | Russian | with the voice owner’s consent | — |
| Piper “Homer” (medium) — the voice on nodes without an NVIDIA GPU | Russian | our own: the same “Homer” voice with its owner’s consent; fine-tuned from en_US-john (MIT) | homer.varyen.net/node/voices |
How Piper “Homer” was made: a corpus of 5,500 phrases — public-domain prose plus speaker phrases — was read by the built-in “Homer” voice, and a Piper voice starting from en_US-john-medium (MIT, LibriVox recordings) was fine-tuned on it. Training used piper1-gpl (GPL-3.0); that code does not ship, the node runs the prebuilt Piper.
Russian voices from piper-voices
The ru_RU voices from piper-voices (dmitri, denis, irina, ruslan) do not ship with Homer 1.0: dmitri and denis are fine-tuned from the lessac voice, whose dataset is for research only; ruslan is CC BY-NC-SA (non-commercial); irina has no stated license. If you already run your own speaches server and the home uses one of them, it keeps speaking with it.
Voice recognition
The models Homer uses to recognize who is speaking are the stock biometrics from Yandex speaker firmware: ECAPA-TDNN for the voice print and gender and age classifiers. These are vendor files: they are not part of the Homer package and the node does not download them; you take them from your own speaker and put them into speaker-home\voiceid. Without them voice recognition stays silent, everything else works.
Node language models
| Model | License | Source |
|---|---|---|
| Qwen3 4B Instruct, Qwen3.5 9B, Qwen3.6 35B A3B (GGUF) | Apache 2.0 | unsloth |
| GigaChat 3.1 10B A1.8B (GGUF) | MIT | ai-sage |