Veridity — on-device files

Files that the Veridity Android app downloads after install. The app checks every file's size and SHA-256 against values built into the app, so a file here is used only if it is exactly the file listed below. SHA256SUMS lists all of them. speech-delivery-manifest.json lists every file inside the two speech ZIPs with its size, SHA-256 and source.

Each file keeps its own licence. They are listed per file below and in NOTICE.

gemma-4-E2B-it-mind-fp32.litertlm — her mind (Apache-2.0)

Google's Gemma 4 E2B instruction-tuned model in LiteRT-LM format.

• Source: gemma-4-E2B-it.litertlm from litert-community/gemma-4-E2B-it-litert-lm (Apache-2.0). Source file SHA-256: 181938105e0eefd105961417e8da75903eacda102c4fce9ce90f50b97139a63c (2,588,147,712 bytes).
• Modification: exactly two bytes changed, at file offsets 298-299. The model's metadata string fp16 becomes fp32, so the phone's GPU runs the model's activations at fp32 precision. The weights are unchanged.
• This file: 2,588,147,712 bytes.
• Licence: Apache License 2.0 (LICENSE).

veridity-hearing-moonshine-base-2026-10-23.zip — her hearing (MIT)

The Moonshine Base English speech-recognition model by Useful Sensors, Inc. (dba Moonshine AI), as converted for sherpa-onnx.

• Source: csukuangfj2/sherpa-onnx-moonshine-base-en-quantized-2026-02-27 at commit 8f4d6c58c03d40bcea40043bb7120a878f2bbef6. Original model and code: moonshine-ai/moonshine.
• Contents: that commit's encoder_model.ort, decoder_model_merged.ort, tokens.txt and LICENSE, byte for byte (unchanged), plus four small Veridity setting files: ears_model.txt, ears_guard.txt (silero), ears_threads.txt (4) and ears_pause_ms.txt (400).
• Licence: MIT License. Copyright (c) 2025 Useful Sensors, Inc. (dba Moonshine AI). The full licence is inside the ZIP. Only the English model is included; the upstream licence's non-commercial terms for other languages do not apply to it.

veridity-voice-p303-trim148-2026-10-23.zip — her voice (CC BY 4.0)

The Pocket TTS text-to-speech model by Kyutai, in the ONNX export by KevinAHM, with Veridity's voice.

• Model source: KevinAHM/pocket-tts-onnx at commit 58a6d00cf13d239b6748cb0769f35c580a8f606c, derived from kyutai/pocket-tts. The fp32 english_2026-04 graphs flow_lm_main.onnx, flow_lm_flow.onnx, mimi_decoder.onnx and text_conditioner.onnx and their bundle.json are byte for byte that commit's files (unchanged). pocket_spm_tokenizer.onnx is a SentencePiece tokenizer graph built for the app's runtime.
• Voice: veridity_voice_mesh20_flat.safetensors is a Pocket TTS voice state made only from two recordings of speaker p303 of the CSTR VCTK Corpus (p303_146_mic1.flac, p303_023_mic1.flac). Both were raised by 2 dB and encoded with Pocket TTS. Parts of the two encodings were combined and the result flattened for the phone. The final silent positions (148 of 162 kept) were then removed. The file name is historical. The voice contains no other speaker.
• Settings files: voice_eos.txt 0.0, voice_gain_db.txt 3.0, voice_rate.txt 0.985, voice_temp.txt 0.70, voice_tilt_db.txt 0, warmth_db.txt 2.5, and the marker file precision_fp32.
• Licence: Creative Commons Attribution 4.0 International (CC BY 4.0), as Pocket TTS and the VCTK Corpus. Changes were made, as described above. Kyutai, KevinAHM, the University of Edinburgh and the VCTK speakers do not endorse Veridity.
• Corpus credit: Yamagishi, Junichi; Veaux, Christophe; MacDonald, Kirsten. (2019). CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit (version 0.92), [sound]. University of Edinburgh, The Centre for Speech Technology Research (CSTR). https://doi.org/10.7488/ds/2645
• Use terms carried forward from the model's source: the voice must not be used for anything illegal, harmful, deceptive or fraudulent. This includes impersonating or cloning a real person's voice without their consent, misinformation, fraudulent calls, presenting generated speech as a real recording of a real person, and harassing, hateful or privacy-invasive content. See the Prohibited Use section of the source model card.