Bundled models and licenses
Ambrogio’s source code is released under the MIT license. The models it ships with are separate works with their own licenses.
Backend
Section titled “Backend”| Model | Use | Location | License |
|---|---|---|---|
| Piper voice Paola (medium), sherpa-onnx conversion | Speech synthesis | /opt/ambrogio/voices/vits-piper-it_IT-paola-medium in the image |
MIT, from rhasspy/piper-voices; converted by k2-fsa |
| Piper voice Riccardo (x_low), sherpa-onnx conversion | Speech synthesis | /opt/ambrogio/voices/vits-piper-it_IT-riccardo-x_low in the image |
MIT, as above |
| Piper voice Kristin (medium), sherpa-onnx conversion | English speech synthesis | /opt/ambrogio/voices/vits-piper-en_US-kristin-medium in the image |
MIT, as above |
| Nemotron 3.5 ASR streaming 0.6B by NVIDIA, 560 ms chunks, multilingual, sherpa-onnx conversion | Italian and English streaming speech recognition | /opt/ambrogio/models/sherpa-onnx-nemotron-3.5-asr-streaming-0.6b-560ms-int8-2026-06-11 in the image: encoder.int8.onnx, decoder.int8.onnx, joiner.int8.onnx, tokens.txt, LICENSE |
OpenMDW 1.1, included LICENSE |
The voices come from the sherpa-onnx tts-models release; the Nemotron model comes from the archive sherpa-onnx-nemotron-3.5-asr-streaming-0.6b-560ms-int8-2026-06-11.tar.bz2 of the sherpa-onnx asr-models release. The image keeps only the int8 encoder, decoder and joiner of Nemotron and its tokens.txt, and leaves out the archive’s README.md and test recordings. OpenMDW 1.1 requires a distribution to keep a copy of the license and the notices of origin: the image adds the license text as LICENSE in the model folder, and the backend repository’s voices/README.md records the origin. The build verifies every archive’s SHA-256 checksum before extraction. The backend repository’s voices/README.md lists all four checksums, for example c6bf5e0df765f9d5b43bc9e0536d4b4b3e7d40bdf5ecf13e45f134c51c05ae3a for the Nemotron archive. You may still mount other models; check their licenses separately. See Configure voice.
Satellite
Section titled “Satellite”| Model | Use | Location | License |
|---|---|---|---|
sherpa-onnx keyword spotting zipformer GigaSpeech 3.3M (int8 encoder, decoder and joiner, tokens.txt, bpe.model) |
Wake word | models/kws/, /app/models/kws/ in the image |
Apache-2.0, models/kws/LICENSE |
Silero VAD silero_vad.onnx |
Detecting when you speak | models/vad/, /app/models/vad/ in the image |
MIT, models/vad/LICENSE, from snakers4/silero-vad |
The keyword spotting model comes from the archive sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.tar.bz2 of the sherpa-onnx kws-models release. It has about 3.3 million parameters and was trained on GigaSpeech, a collection of English speech; only its int8 files are shipped. Silero VAD comes from the sherpa-onnx asr-models release. The files are copied unchanged, and the satellite repository’s models/README.md lists their SHA-256 checksums. Both licenses allow redistribution, commercial use included. A keyword spotting model you mount yourself has its own license; see Change the wake word.
The satellite ships no recordings: its beep and its error tone are generated by the program.