AswanthCManoj/moonshine-eou-detector

Moonshine EOU Detector

End-of-utterance (EOU) confirmation gate for streaming voice pipelines — runs alongside Silero VAD to decide when a caller's turn is truly finished.

All-ONNX inference, no PyTorch at runtime.

Files

File	Size	Description
moonshine_enc_fp32.onnx	0.9 MB	Moonshine-tiny encoder graph (7.7M params, 288D hidden)
moonshine_enc_fp32.onnx.data	30 MB	Encoder weights (external data)
eou_head.onnx	2.2 MB	Trained binary classification head
silero_vad.onnx	2.2 MB	Silero VAD v5 (bundled for convenience)

Usage

import numpy as np
import onnxruntime as ort

# Load sessions
enc = ort.InferenceSession("moonshine_enc_fp32.onnx")
head = ort.InferenceSession("eou_head.onnx")

# Predict P(turn complete) from 16 kHz float32 audio
wav = np.zeros(16000, dtype=np.float32).reshape(1, -1)  # 1 second
features = enc.run(None, {"input_values": wav})[0]
frame_len = np.array([features.shape[1]], dtype=np.int64)
logit = head.run(None, {"encoder_features": features, "frame_lengths": frame_len})[0]
prob = float(1.0 / (1.0 + np.exp(-logit[0])))

EOU Schedule

The model is consulted at up to 4 checkpoints during caller silence, each with a relaxing gate:

Silence	Threshold	Rationale
96 ms	0.85	Very confident — fast commit
192 ms	0.65	Moderate confidence
320 ms	model default (0.50)	Calibrated threshold
480 ms	0.40	Lenient — long pause likely means done
576 ms	(timeout)	Hard backstop, commit regardless

Silero VAD ONNX — Context Prefix

The bundled silero_vad.onnx requires 64 samples of context prepended to each 512-sample window (576 total input). Without this, the model produces near-zero probabilities and speech detection silently fails.

The official Silero OnnxWrapper and PyTorch model handle this internally. Custom ONNX code must replicate it:

_CONTEXT_SIZE = 64
context = np.zeros(64, dtype=np.float32)

def process_chunk(chunk_512):
    x = np.concatenate([context, chunk_512.astype(np.float32)])
    context[:] = x[-_CONTEXT_SIZE:]
    out, state = session.run(None, {
        "input": x.reshape(1, -1),
        "state": state,
        "sr": np.array(16000, dtype=np.int64),
    })
    return float(out[0, 0])

Architecture

• Encoder: Moonshine-tiny (7.7M params), passthrough feature extractor — feed raw 16 kHz audio directly as (1, seq) float32
• Head: Linear projection from pooled encoder features to binary logit, trained on turn-boundary data
• VAD: Silero v5 ONNX with per-session state (2, 1, 128) (~1 KB) + 64-sample context buffer