Add OpenWakeWord Alexa F32 GGUF

.gitattributes

@@ -155,3 +155,4 @@ Voxtral-4B-TTS-2603-GGUF/voxtral-4b-tts-2603-bf16.gguf filter=lfs diff=lfs merge
 Voxtral-4B-TTS-2603-GGUF/voxtral-4b-tts-2603-q8_0.gguf filter=lfs diff=lfs merge=lfs -text
 Plapre-3-GGUF/plapre-3-bf16.gguf filter=lfs diff=lfs merge=lfs -text
 Plapre-3-GGUF/plapre-3-f32.gguf filter=lfs diff=lfs merge=lfs -text
+OpenWakeWord-GGUF/open-wake-word-alexa-f32.gguf filter=lfs diff=lfs merge=lfs -text

OpenWakeWord-GGUF/README.md

@@ -0,0 +1,57 @@
+---
+license: cc-by-nc-sa-4.0
+language:
+- en
+pipeline_tag: audio-classification
+tags:
+- audio.cpp
+- gguf
+- wake-word-detection
+---
+
+# OpenWakeWord Alexa GGUF
+
+F32 conversion of the official [OpenWakeWord v0.5.1 release](https://github.com/dscripka/openWakeWord/releases/tag/v0.5.1)
+frontend, speech embedding model, and Alexa v0.1 classifier.
+
+`open-wake-word-alexa-f32.gguf` is a single-file package containing the weights,
+frontend coefficients, configuration, fixed initial feature history, and the
+audio.cpp v1 model spec. No Python, ONNX Runtime, or additional weights are needed
+for C++ inference. This classifier detects `Alexa`; it is not a general keyword
+spotter.
+
+## Usage
+
+Input must be mono PCM audio at 16 kHz.
+
+```bash
+audiocpp_cli --family open_wake_word --task wake --backend cpu \
+  --model OpenWakeWord-GGUF/open-wake-word-alexa-f32.gguf \
+  --audio input.wav --request-option threshold=0.5 \
+  --segments-out detections.json --log --metrics
+```
+
+Add `--mode streaming` and omit `--metrics` for streaming input. The default
+threshold is 0.5. Detections are rising threshold crossings labelled `Alexa`,
+reported at an 80 ms cadence; the first five predictions are suppressed.
+
+See the [model guide](https://github.com/0xShug0/audio.cpp/blob/main/docs/models/open_wake_word.md)
+for conversion and component execution-plan controls.
+
+## Validation
+
+CPU, CUDA, and Vulkan detected the official Alexa recording at samples
+7680-8960 (0.48-0.56 seconds), with confidence 0.999996. Offline and streaming
+detection JSONs matched across all three backends. A separate 14-second speech
+recording without the wake word produced no detections on any backend.
+
+These are correctness checks, not a general false-accept/false-reject benchmark.
+
+## License
+
+The upstream project declares its pretrained models under
+[CC-BY-NC-SA-4.0](https://github.com/dscripka/openWakeWord#license); the GGUF
+retains those terms and is not licensed for commercial use. Upstream source code
+is separately Apache-2.0. Attribution: David Scripka and the OpenWakeWord
+contributors. Conversion packages the pretrained tensors for audio.cpp without
+changing their precision.

OpenWakeWord-GGUF/open-wake-word-alexa-f32.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:649c2b6e10f9ffc23923250931e58edfe9c42d293660e02b32aba09383d9a509
+size 2260928