@neuvybe/jarvis 0.46.2 → 0.46.3

package/dist/assets/LICENSE-silero-vad

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020-present Silero Team
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/dist/assets/LICENSE-smart-turn-v3

@@ -0,0 +1,24 @@
+BSD 2-Clause License
+
+Copyright (c) 2024-2025, Daily
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

package/dist/assets/MODEL.md

@@ -0,0 +1,113 @@
+# Vendored models
+
+This directory holds the third-party ONNX models Jarvis ships and runs itself on the
+shared onnxruntime-web WASM runtime (proposal 0026):
+
+- `smart-turn-v3.1-cpu.onnx` - the prosody endpointer (this section)
+- `silero_vad.onnx` - the speech/no-speech VAD (second section; proposal 0028 R1)
+
+# smart-turn-v3.1-cpu.onnx
+
+The prosody endpointer behind the Natural Turn Controller (proposal 0024 Decision 5).
+It answers "does this intonation sound like a finished thought?" so Jarvis can tell a
+mid-thought pause from a finished turn and not cut the user off.
+
+## What it is
+
+- File: `smart-turn-v3.1-cpu.onnx`
+- Architecture: Whisper-tiny encoder + a linear turn-completion classifier (~8M params)
+- Quantization: int8, CPU build (the `cpu` variant; no GPU/DirectML/CUDA needed)
+- Input: `input_features`, float32 `[batch, 80, 800]` - Whisper-tiny log-mel of an 8 s
+  window (80 mel bins x 800 frames at 16 kHz, hop 160)
+- Output: `logits`, float32 `[batch, 1]`. Despite the name this is ALREADY a probability
+  in `[0, 1]` - the sigmoid is baked into the graph. Read it directly as P(complete); do
+  NOT apply another sigmoid (that double-sigmoid was the voice-deps spike's
+  miscalibration).
+
+## Provenance
+
+- Source: https://huggingface.co/pipecat-ai/smart-turn-v3 (file
+  `smart-turn-v3.1-cpu.onnx`)
+- Upstream project: https://github.com/pipecat-ai/smart-turn
+- Downloaded: 2026-06-09
+- Size: 8,679,180 bytes
+- SHA-256: `fb68d55c2d542ce79e44b12013bfd571e90df8594ab096d757198e851b0c6594`
+
+To re-verify the file, run this from the directory that contains it
+(`packages/core/assets/` in the repo, `dist/assets/` in an installed package):
+
+```
+sha256sum smart-turn-v3.1-cpu.onnx
+```
+
+## License
+
+BSD 2-Clause, Copyright (c) 2024-2025, Daily. The full text is in
+`LICENSE-smart-turn-v3` alongside this file. Redistributed unmodified.
+
+## Calibration
+
+Our TypeScript log-mel front end (`packages/core/src/speech/log-mel.ts`) is a faithful
+port of pipecat's vendored numpy reference. The committed fixtures under
+`packages/core/src/speech/__fixtures__/smart-turn/` pin our features against that
+reference: they match to float32 round-off (~1e-5 worst case; our waveform normalization
+runs in float64 where the reference uses float32). That feature parity is the
+platform-stable calibration anchor. The model output is read directly as P(complete) (no
+double-sigmoid); the int8 model's absolute output is not bit-reproducible across
+ONNX-runtime builds / CPUs, so the exact P is not pinned cross-platform. See the
+calibration tests and `docs/notebook/2026-06-09-smart-turn-calibration.md`.
+
+# silero_vad.onnx
+
+The speech/no-speech classifier (proposal 0028 R1): it replaces the RMS-energy speech
+detection in both the core's Turn Controller (per-frame speech/silence driving turn-end)
+and the CLI's barge-in onset - the macos-capture-rate wrench's named follow-up. Energy
+detects loudness; Silero detects speech.
+
+## What it is
+
+- File: `silero_vad.onnx`
+- Architecture: Silero VAD v5 (STFT front end + conv encoder + recurrent decoder,
+  ~1.8M params, float32)
+- Input: `input`, float32 `[batch, 576]` = 64 samples of CONTEXT (the last 64 samples
+  of the previous window; zeros at stream start) + 512 new samples at 16 kHz (~32 ms
+  per window); `state`, float32 `[2, batch, 128]` (recurrent state carried across
+  windows within a stream); `sr`, int64 = 16000
+- Output: `output`, float32 `[batch, 1]` = P(speech) for the window; `stateN`, float32
+  `[2, batch, 128]` = the next recurrent state
+- IMPORTANT: the graph silently ACCEPTS a context-less `[batch, 512]` input and returns
+  garbage (~0.0006 for speech and silence alike). The official silero-vad OnnxWrapper's
+  64-sample context prepend is load-bearing; `packages/core/src/speech/silero-vad-runner.ts`
+  implements it and documents the verification (2026-06-11).
+
+## Provenance
+
+- Source: copied from decibri's bundled model (`node_modules/decibri/models/silero_vad.onnx`,
+  decibri 4.3.0) - the same artifact decibri runs in its own `vad: 'silero'` mode, so its
+  calibration defaults (threshold 0.5, holdoff 300 ms) carry over
+- Upstream project: https://github.com/snakers4/silero-vad (the v5 `silero_vad.onnx`
+  release artifact decibri redistributes)
+- Copied: 2026-06-11
+- Size: 2,327,524 bytes
+- SHA-256: `1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3`
+
+To re-verify the file, run this from the directory that contains it
+(`packages/core/assets/` in the repo, `dist/assets/` in an installed package):
+
+```
+sha256sum silero_vad.onnx
+```
+
+## License
+
+MIT, Copyright (c) 2020-present Silero Team. The full text is in `LICENSE-silero-vad`
+alongside this file. Redistributed unmodified.
+
+## Calibration
+
+The probability is read directly from `output` (no post-processing). decibri's proven
+defaults are mirrored where the model is consumed: speech threshold 0.5
+(`SileroFrameClassifier` / `SileroBargeInDetector`), onset debounce ~2 windows, and
+re-arm holdoff 300 ms (the barge-in detector). The model is float32 (not quantized), so
+tests assert speech-vs-silence separation on synthesized audio rather than absolute
+probabilities; see `silero-vad-runner.test.ts`.