DhawalM/mira-invocation-audio

Mira invocation audio

An exact snapshot of the active audio dataset used by the Mira Whisper and HuBERT invocation classifiers. It includes 167 derivatives from one real microphone training session.

The label is 1 when the glasses wearer directly addresses the Mira assistant, and 0 for incidental mentions, other conversation, or background audio. Only the wearer's direct address should activate Mira. Synthetic positives assume the synthetic speaker is the wearer; the classifiers do not verify speaker identity.

Files and splits

All 2,787 WAV files are in one flat data/audio/ directory. The unchanged data/manifest.csv is the authoritative manifest. The audio_path field resolves from the repository root. Use the existing split column:

Split	Clips	Positive	Negative
train	2,512	1,079	1,433
validation	157	61	96
test	118	59	59

Every stored input is 30 seconds, mono, 16 kHz. Stored duration is 23.225 hours, including silence, synthetic variants, and repeated source material; this is not 23 hours of independent speech. WAVs total about 3.94 GB in decimal units. splits.csv records counts and stored durations. checkpoints.csv maps the saved models to their exact training runs and validation results. Individual model repositories contain the corresponding run-specific manifest.csv snapshots.

The 510 most recent training windows embed 16 seconds of context after 14 seconds of artificial left padding. Older examples can contain a full 30 seconds of context. This differs from the live demo's uniformly selected 16-second rolling crop, left-padded to 30 seconds and scored every second.

Provenance and evaluation limits

The data combines local Qwen TTS, Kokoro, and Piper generations, prepared acoustic variants, and one real microphone session. Consult engine, model, voice, source_*, and padding fields where populated. Older rows have incomplete provider metadata. Text fields support provenance and labels; model inference uses audio only.

Real microphone labels were assisted by ASR and have not been fully human-verified. Derivatives of the same recording are not independent examples. The exact model input WAVs are included; raw source sessions, intermediate clips, annotation files, and old paths referenced only by provenance columns are not part of this snapshot.

Whisper and expanded HuBERT used 2,512 training and 157 validation clips. The earlier HuBERT real-microphone baseline used 2,002 training and the same 157 validation clips. Synthetic validation selected both checkpoints and thresholds, so its metrics are not independent real-world results or false activations per hour. The test split was not used for Whisper training or checkpoint/threshold selection; an earlier HuBERT comparison evaluated it. Preserve the recorded split assignments.

Download into the training checkout

Run from the root of the training code checkout:

hf download DhawalM/mira-invocation-audio --repo-type dataset --include 'data/*' --local-dir .

The single CSV manifest layout is retained as a file archive. Dataset Viewer is disabled to avoid automatic file-based split inference; load rows with the project loader, which explicitly filters the split column.

Models: Whisper and HuBERT.

This handoff does not assign a new redistribution license to the recordings or synthetic voice outputs. Their underlying source terms remain applicable.