ECHO-1 on-device model packs

Optional downloads for the ECHO-1 offline survival assistant. Everything runs on the device, without a network connection, on iOS, Android, macOS and Linux (including Raspberry Pi). The app reads manifest.json, downloads a pack's files from a fixed commit and checks each file's SHA-256 before using it.

Pack	Files	Size	What it is	Source	License
hearing/	small-encoder.int8.onnx, small-decoder.int8.onnx, small-tokens.txt	375 MB	OpenAI Whisper small, 8-bit ONNX export for sherpa-onnx; speech to text in 99 languages	csukuangfj/sherpa-onnx-whisper-small	MIT (OpenAI Whisper)
hearing/	silero_vad.onnx	2 MB	Silero voice activity detector	snakers4/silero-vad via csukuangfj/vad	MIT
speaking/	*.onnx, tts.json, unicode_indexer.bin, voice.bin	145 MB	Supertonic 3 text to speech, 8-bit build for sherpa-onnx, 31 languages	k2-fsa/sherpa-onnx tts-models, model by Supertone Inc.	Code MIT (speaking/LICENSE); model weights under Supertone's OpenRAIL-M license, use restrictions apply
vision/	gemma-4-E2B-vision-q8_0.gguf	197 MB	Image encoder of Google Gemma 4 E2B, for llama.cpp mtmd. Made from Google's gemma-4-E2B-it-mmproj.gguf by removing the audio encoder and storing the vision matrices as Q8_0 with scripts/make_vision_mmproj.py	google/gemma-4-E2B-it-qat-q4_0-gguf	Apache 2.0, Gemma 4 license

The vision file only works with Gemma 4 E2B (gemma-4-E2B_q4_0-it.gguf).

Changes made

• vision/: Google's original file holds an image encoder (367 MB) and an audio encoder (612 MB) in BF16. The audio encoder and its projector were removed; the image encoder's block matrices are Q8_0, the position table F16, the rest F32. In greedy tests it gives the same output as the original file.
• hearing/, speaking/: unchanged copies of the upstream files.

All rights to the models remain with their authors. Use of each model is subject to its license above.