rhasspy/voice-labs-models

Voice Labs models

The models the Voice Labs app downloads the first time it starts, as it loads them (Parakeet, GLiNER), and those its image's build fetches (kaldi/, micro-wake-word/).

Folder	What	From	License
parakeet-tdt_ctc-110m-en-int8/	NVIDIA's Parakeet TDT-CTC 110M, English: the transducer (encoder, decoder, joiner, tokens.txt) with its encoder's MatMul weights quantized to int8 (onnxruntime's dynamic quantization), and the CTC branch's head (ctc_head.npz), so that one encoder gives both the CTC log-probabilities commands are matched on and the transcripts	nvidia/parakeet-tdt_ctc-110m, as exported by csukuangfj/sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000 (e9bea5a) and csukuangfj/sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000 (3af92f1)	CC-BY-4.0, © NVIDIA Corporation
gliner2.5-small-v1-onnx/	fastino's GLiNER 2.5 small for the app's command fallback: ONNX networks with the encoder's layers in 8 bits and its embedding in 4 (MatMulNBits, block 32), the classifier's weights and the tokenizer (gliner2_onnx.json says how)	fastino/gliner2.5-small-v1	Apache-2.0
kaldi/	The wake word verifier's Kaldi. en_US-rhasspy.tar.gz: the U.S. English nnet3 acoustic model and IPA lexicon of Rhasspy's Kaldi profile, without its G2P model and language model (the verifier never guesses a pronunciation). tools-x86_64.tar.gz and tools-aarch64.tar.gz: Kaldi's and OpenFst's binaries, libraries, headers and utils//steps/ scripts, which the image's build compiles the verifier's resident decoder against	rhasspy/en-us_kaldi-rhasspy; Kaldi and OpenFst 1.6.7 -- x86-64 built from source, aarch64 the Kaldi of 2021 that Rhasspy shipped (kaldi-manylinux)	MIT (the model); Apache-2.0 (Kaldi, OpenFst)
micro-wake-word/	The microWakeWord models the app pushes to satellites, with their settings: okay_nabu (multilingual) and stop as Kevin Ahrendt released them, vad, hey_jarvis and hey_mycroft as ESPHome publishes them	kahrendt/microWakeWord releases okay_nabu_20241226.3 and stop; esphome/micro-wake-word-models (models/v2/)	Apache-2.0

The app also downloads EmbeddingGemma 2 (onnx/model_q4.onnx and its tokenizer) from onnx-community/embeddinggemma-2-ONNX, and its image's build ReDimNet-b2 (speaker identification) from OpenVoiceOS/redimnet-b2-vox2-onnx.