Documentation
¶
Overview ¶
Package wakewordcatalog is the kernel-neutral source of truth for the wake phrases SpeechKit trains and serves, across BOTH engine families:
- openWakeWord (ONNX): consumed host-side by the desktop reference app and the box companion. Trained via tools/wakeword-training (livekit-wakeword).
- microWakeWord (TFLite + JSON manifest): consumed ON-DEVICE by ESPHome voice satellites and the Kombify-Box firmware (esp-tflite-micro).
One phrase (e.g. "hey_kombify") is therefore describable once here and served to either family. The package is pure data + accessors: no HTTP, no config, no build tag, so it links into the kernel, the Linux server adapter, and tooling alike.
microWakeWord artifacts carry an Available flag. It is false until the microWakeWord candidate pipeline (tools/wakeword-training/microwakeword) has produced a .tflite, all manifest params have independent measurements, the exact bytes have passed physical validation, and publication is verified; the ONNX artifacts are already trained and published.
Index ¶
Constants ¶
const ( MicroWakeWordType = "micro" MicroWakeWordManifestVersion = 2 DefaultMinimumESPHomeVersion = "2024.7.0" // MicroWakeWordFeatureStepSize is fixed by the micro_wake_word streaming // frontend (10 ms slices); it is not a per-model tunable. MicroWakeWordFeatureStepSize = 10 )
microWakeWord manifest constants (ESPHome micro_wake_word v2 schema).
const License = "apache-2.0"
License is the SPDX-ish identifier for every SpeechKit-trained wake model. All phrases are trained from scratch (see tools/wakeword-training/README.md) specifically so the result is permissively licensed for OSS + commercial use.
Variables ¶
var ( URL: hfBase + "melspectrogram.onnx", SHA256: "ba2b0e0f8b7b875369a2c89cb13360ff53bac436f2895cced9f479fa65eb176f", SizeBytes: 1_087_958, } URL: hfBase + "embedding_model.onnx", SHA256: "70d164290c1d095d1d4ee149bc5e00543250a7316b59f31d056cff7bd3075c1f", SizeBytes: 1_326_578, } )
SharedMelspec and SharedEmbedding are the phrase-independent openWakeWord frontend models. Every openWakeWord phrase requires both, alongside its own per-phrase head. microWakeWord models do NOT use these (the TFLite model is self-contained).
Functions ¶
This section is empty.
Types ¶
type FileArtifact ¶
FileArtifact is a downloadable model file: a stable URL plus an integrity hash and size. SHA256 is lower-case hex of the raw bytes.
type MicroWakeWordArtifact ¶
type MicroWakeWordArtifact struct {
Available bool
File FileArtifact
Params MicroWakeWordParams
}
MicroWakeWordArtifact describes the on-device TFLite model plus the manifest parameters ESPHome / esp-tflite-micro need to run it. Available is false until governed publication has been independently verified at the serving origin.
type MicroWakeWordParams ¶
type MicroWakeWordParams struct {
ProbabilityCutoff float64
SlidingWindowSize int
FeatureStepSize int
TensorArenaSize int
MinimumESPHomeVersion string
}
MicroWakeWordParams mirrors the "micro" object of an ESPHome microWakeWord v2 model manifest. ProbabilityCutoff, SlidingWindowSize and TensorArenaSize are evidence-bound before publication; FeatureStepSize is fixed (see the const above). TensorArenaSize comes from the target runtime rather than the model trainer.
type Model ¶
type Model struct {
ID string // stable slug, e.g. "hey_kombify"
WakeWord string // spoken phrase, e.g. "Hey Kombify"
DisplayName string // human label for dashboards
Description string // short tuning/pronunciation note
TrainedLanguages []string // BCP-47-ish language tags the model was trained on
OpenWakeWord OpenWakeWordArtifact
MicroWakeWord MicroWakeWordArtifact
}
Model is one wake phrase available across both engine families.
func All ¶
func All() []Model
All returns a copy of the model registry. Callers must not mutate the shared slice contents; the returned slice header is safe to range over.
func (Model) HasMicroWakeWord ¶
HasMicroWakeWord reports whether a published microWakeWord TFLite exists for this phrase (Available AND a real file URL).
func (Model) HasOpenWakeWord ¶
HasOpenWakeWord reports whether the model ships a published openWakeWord head.
type OpenWakeWordArtifact ¶
type OpenWakeWordArtifact struct {
File FileArtifact
RecommendedThreshold float64
}
OpenWakeWordArtifact is the per-phrase conv-attention ONNX head. The shared melspectrogram + embedding frontends every phrase depends on are SharedMelspec and SharedEmbedding below.
Source Files