jbuaba/attune-cadence

Attune Cadence

Attune Cadence is a compact English speech-perception system that preserves both what was said and how it sounded. For a 0.5–30 second voice recording, it returns a transcript, word timing, localized vocal events, phrase-level perceived affect, confidence, and abstention in one time-aligned JSON result.

Cadence reports observable vocal evidence and uncertain listener perception; it does not claim to determine a speaker's internal emotional state. The JSON schema is authoritative. XML and bracketed transcripts are deterministic views generated from the validated result.

Cadence v0.13

v0.13 is the current evaluation model. It keeps the 241,904,650-parameter Cadence INT8 graph for transcription, timing, and vocal events, and adds a 57,937,364-parameter affect branch. The complete model uses 299,842,014 active parameters.

The model returns transcripts, word timing, conservative vocal events, phrase-level perceived affect, and abstention. Vocal styles and dimensional valence, arousal, and dominance remain disabled until they pass the project's evaluation gates.

Run v0.13 with the root attune-cadence-242m-int8.onnx graph, candidate-v0.13/affect-int8.pt, and candidate-v0.13/fused-calibration.json. The candidate directory also includes its manifest, artifact hashes, evaluation reports, and full results.

This package is currently private. v0.13 is intended for consented research and evaluation, not as a verified emotion detector or for automated high-stakes decisions.

Scope

• English, single-speaker clips or user turns.
• 16 kHz mono PCM16 WAV input, 0.5–30 seconds.
• CTC transcript and approximate word timings.
• Calibrated laugh, sigh, cough, throat_clear, and sneeze events.
• One utterance-level perceived-affect distribution, plus experimental fixed windows for recordings longer than four seconds.
• Abstention and OOD probability.
• Vocal styles and V/A/D disabled in v0.1.

For ordinary speech without supported paralinguistic evidence, a readable view can be:

[affect uncertain] I hate you, I hate you so much—never call me again.

Four neutral system voices reading that exact sentence are the regression for the false-tag failure that prompted this candidate. FP32 and INT8 transcribe all four correctly, produce no events or styles, and abstain on affect.

Architecture and files

The SenseVoice-derived ONNX graph has 241,904,650 parameters. A separate 30,726-parameter NumPy event head consumes a padding-safe acoustic embedding from the same graph. The unique total is 241,935,376 parameters.

• attune-cadence-242m-int8.onnx — recommended 500 MB deployment graph.
• attune-cadence-242m-fp.onnx — 970 MB full-precision reference.
• event-head.npz — calibrated event head; no PyTorch dependency.
• calibration.json — affect and OOD calibration.
• attune-output-v2.0.schema.json — authoritative output contract.
• evidence/ — evaluation and deployment reports.
• release-gates.json — executable publication decision.

The Python runtime also needs the official SenseVoiceSmall frontend and decoder assets obtained separately under the upstream agreement.

Measured results

Metric	Full precision	INT8
VocalSound event macro-F1, 80 opened clips	0.814	0.826
External OOD false-positive rate, 160 clips	0.0125	0.0125
Exact hostile-lexical transcripts	4/4	4/4
False events/styles on those controls	0	0

The retained full-precision affect graph reaches macro-F1 0.6643 on development, 0.6336 on an opened regression set, and 0.3853 on external RAVDESS. INT8 reaches 0.3803 on the same 480 clips, an absolute loss of 0.0051. Both external results are below the project's 0.40 public-release target. No public affect-generalization claim is supported.

An internal transition regression concatenated one RAVDESS joy clip and one distress clip. The compact head classified both component windows as fear, whereas the local emotion2vec+ teacher classified the original clips correctly. This regression is not a general benchmark, but it is a concrete reason to treat the current affect windows as experimental.

The event sets are source-labelled, partly isolated sounds rather than reviewed natural inline events. The four lexical controls are synthetic neutral speech. These opened results do not replace a sealed human evaluation.

Intended use and safety

Use Cadence only as uncertain supplementary evidence in consented speech research, accessibility experiments, and low-stakes conversational interfaces. Do not use it for covert monitoring, diagnosis, deception detection, protected- trait inference, speaker identification, or automated high-stakes decisions. Do not state that a speaker is an emotion based on this output.

No controlled downstream human study has established that this representation improves conversational responses.

Licence and attribution

Cadence is derived from SenseVoiceSmall by FunASR/FunAudioLLM. Its weights and derivatives are governed by the FunASR Model Open Source License Agreement v1.1, separately from the MIT-licensed Attune source code. Dataset terms and the current redistribution decision are included with the project. A private Hugging Face upload is not public-release clearance.