Insta360 Wave — AI Model Extraction Guide

How to extract all AI/ML models from the Insta360 Wave device.

---

Overview

The Wave has models in two locations:

Location	Partition	Access via USB File Transfer	Access via ADB
/usr/bin/tea_models/	rootfs (SquashFS)	No	Yes
/model/	Separate ext4 (4.1 GB)	No	Yes
/usr/lib/s2t_data/	rootfs (SquashFS)	No	Yes

USB File Transfer mode only exposes /mnt/udisk/ (the user storage partition with recordings). The model and rootfs partitions require root access via ADB.

---

Step-by-Step Extraction

Prerequisites

brew install android-platform-tools
pip3 install hidapi  # or: pip3 install --break-system-packages hidapi

1. Enable ADB

Use the checked-in helper script from the repository root:

If you are using the standalone copy instead of the checked-in helper:

python3 enable_adb.py
sleep 3
adb devices  # Should show PSP_<serial>

If you need a standalone copy outside this repo, this is the helper script contents:

#!/usr/bin/env python3
import hid

h = hid.device()
h.open(0x18F0, 0x4E40)  # Insta360 Wave
h.set_nonblocking(1)
while h.read(64): pass  # drain

pkt = bytearray(64)
pkt[0] = 0x03  # Report ID
pkt[1] = 0xF0  # HID_MSG_TYPE_OF_SWT_FACTORY
pkt[2] = 0x00
pkt[4] = 0xF0
pkt[5] = 0xC0  # ADB enable
h.write(bytes(pkt))
h.close()
print("ADB enabled. Run: adb devices")

python3 ./waved/scripts/enable_adb.py
sleep 3
adb devices  # Should show PSP_<serial>

If the device rebooted or switched USB modes, run the helper again before pulling files.

2. Pull All Models

# Create output directory
mkdir -p extracted_models

# ASR models from /model partition (4.1 GB, ~10 min over USB 2.0)
adb pull /model/ extracted_models/model/

# TEA SDK audio models from rootfs (75.5 MB)
adb pull /usr/bin/tea_models/ extracted_models/tea_models/

# Chinese conversion data (1 MB)
adb pull /usr/lib/s2t_data/ extracted_models/s2t_data/

# Inference runtime libraries
mkdir -p extracted_models/runtime
adb pull /usr/lib/libonnxruntime.so.1.15.1 extracted_models/runtime/
adb pull /usr/lib/libnnsdk.so extracted_models/runtime/
adb pull /usr/lib/libasr_psp.so extracted_models/runtime/
adb pull /usr/lib/libasr_reverb.so extracted_models/runtime/
adb pull /usr/lib/libTEAProcessSdk.so extracted_models/runtime/

3. Verify

# Check file count (should be 173 for /model, 15 for tea_models)
find extracted_models/model/ -type f | wc -l
find extracted_models/tea_models/ -type f | wc -l

# Check model version
cat extracted_models/model/model_version.txt
# Expected: 20250613_V2.0.0.1

4. Restore Normal USB Mode

#!/usr/bin/env python3
import hid
h = hid.device()
h.open(0x18F0, 0x4E40)
h.set_nonblocking(1)
while h.read(64): pass
pkt = bytearray(64)
pkt[0] = 0x03
pkt[1] = 0xF0
pkt[2] = 0x00
pkt[4] = 0xF0
pkt[5] = 0xE0  # User mode (restore normal USB)
h.write(bytes(pkt))
h.close()
print("Normal USB mode restored.")

---

Extracted Model Inventory

Model version: 20250613_V2.0.0.1

ASR Models (/model/) — 4.1 GB total

Chinese ASR (1.2 GB)

Directory	Files	Format	Architecture
asr_ch/offline_sensevoice/	17 .nb + config	VeriSilicon NBG	SenseVoice (Alibaba)
asr_ch/online_zipformer/	1 .nb + tokenizer	VeriSilicon NBG	Zipformer (k2-fsa)

The SenseVoice model is split into 15 sub-graph files (svs6_sp_* and svs12_sp_*) for the NPU — the 6 and 12 variants likely correspond to different batch sizes or chunk lengths.

English ASR (1.1 GB)

Directory	Files	Format	Architecture
asr_en/reverb-asr-10s/	38 .nb + config	VeriSilicon NBG	Reverb-ASR (WeNet)
asr_en/reverb_asr_online/	2 .nb + 1 .onnx + config	Mixed	Reverb-ASR online

The English model has both 5-second and 10-second chunk variants (reverb_5s_split_* and reverb_10s_*), each split into ~19 sub-graphs for NPU execution. The online model has a ctc.onnx (9.8 MB) that runs on CPU via ONNX Runtime and an encoder.nb (177 MB) that runs on NPU.

German ASR (603 MB)

Directory	Files	Format	Architecture
asr_de/de_offline/	28 .nb	VeriSilicon NBG	WeNet
asr_de/de_online/	2 .nb	VeriSilicon NBG	WeNet

Split into encoder_split_6_* (6-chunk) and encoder_split_12_* (12-chunk) variants.

French ASR (604 MB)

Directory	Files	Format	Architecture
asr_fr/fr_offline/	28 .nb	VeriSilicon NBG	WeNet
asr_fr/fr_online/	2 .nb	VeriSilicon NBG	WeNet

Same split structure as German.

Japanese ASR (75 MB)

Directory	Files	Format	Architecture
asr_ja/ja_online/	2 .nb	VeriSilicon NBG	WeNet

Online streaming only — no offline model. Smallest language model.

Speaker Diarization (17 MB)

File	Format	Architecture
diarization/campp_118_cn_en_fp16.nb	VeriSilicon NBG (FP16)	CAM++ (Alibaba 3D-Speaker)

Speaker embedding extraction. FP16 quantized for NPU. Supports Chinese + English speakers.

Voice Activity Detection (2.2 MB)

File	Format	Architecture
vad/silero_vad_v5.onnx	ONNX	Silero VAD v5

The only standard ONNX model — runs on CPU via ONNX Runtime. Open source: https://github.com/snakers4/silero-vad

Punctuation Models (550 MB)

Directory	File	Size	Format	Architecture
punc_model/	model_quant.onnx	269 MB	ONNX	CT-Transformer (online punctuation)
punc_mino/	model_quantized.onnx	266 MB	ONNX	PuncMino (offline punctuation)

Both are standard ONNX models running on CPU. They’re the largest ONNX files.

TEA SDK Audio Models (/usr/bin/tea_models/) — 75.5 MB

File	Size	Format	Purpose
rafa-best_model-11000_model.bin	28 MB	TNNE	Neural noise suppression
insta360_psp_mchpre_chimera_model.bin	13 MB	TNNE	Multi-channel source separation
insta360_nnpost_solo_steedy.bin	6.1 MB	TNNE	Solo speaker processing
insta360_az_taer_*_low_rotate_model.bin	5.1 MB × 3	Binary	Beamforming (low freq)
insta360_az_taer_*_high_rotate_model.bin	2.5 MB × 3	Binary	Beamforming (high freq)
insta360_psp_pre_chimera_model.bin	2.5 MB	TNNE	Source separation pre-mask
premask_reverb_best_model-220000_model.bin	2.5 MB	TNNE	Dereverberation mask
nnec_two_stage_32k.bin	628 KB	TNNE	Neural echo cancellation
insta360_az_taer_*_vad_rotate_model.bin	4.3 KB × 3	Binary	Beamforming VAD

---

Model Formats

VeriSilicon NBG (.nb)

Pre-compiled binary graphs for the Amlogic NPU (VeriSilicon Vivante). Cannot be converted back to ONNX — they are hardware-specific compiled binaries containing:

• Layer operations mapped to NPU hardware units
• Weight data in quantized format (FP16 or INT8)
• Memory allocation plans for NPU SRAM
• DMA transfer schedules

To use these models, you need the same Amlogic A5 NPU hardware + libnnsdk.so ADLA runtime.

To recreate them, you need the original ONNX models + VeriSilicon Acuity Toolkit (NDA required from Amlogic/VeriSilicon).

ONNX (.onnx)

Standard Open Neural Network Exchange format. Can be loaded by ONNX Runtime on any platform.

Files in this format:

• vad/silero_vad_v5.onnx (2.2 MB)
• asr_en/reverb_asr_online/ctc.onnx (9.8 MB)
• punc_model/model_quant.onnx (269 MB)
• punc_mino/model_quantized.onnx (266 MB)

These are portable — you can run them on any machine with ONNX Runtime.

TNNE (.bin with tnne header)

Roomix TEA SDK’s custom neural network format. PyTorch-trained models compiled for their SDK.

Header format:

tnne\r\n
num_layers: <N>
training: pytorch

<layer definitions>

Versioned Binary (.bin with version:0.1 header)

Beamforming coefficient files.

Header format:

version:0.1
<timestamp>
<binary coefficient data>

---

Using the Extracted ONNX Models

Silero VAD

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("extracted_models/model/vad/silero_vad_v5.onnx")

# 16kHz audio, 512-sample chunks
audio_chunk = np.random.randn(1, 512).astype(np.float32)
h = np.zeros((2, 1, 64)).astype(np.float32)
c = np.zeros((2, 1, 64)).astype(np.float32)
sr = np.array([16000], dtype=np.int64)

output, h_out, c_out = session.run(None, {
    "input": audio_chunk, "h": h, "c": c, "sr": sr
})
speech_prob = output[0][0]
print(f"Speech probability: {speech_prob:.3f}")

Punctuation Model

import onnxruntime as ort

session = ort.InferenceSession("extracted_models/model/punc_model/model_quant.onnx")
# Check input/output names
for inp in session.get_inputs():
    print(f"Input: {inp.name} shape={inp.shape} type={inp.type}")
for out in session.get_outputs():
    print(f"Output: {out.name} shape={out.shape} type={out.type}")

---

Automation Script

Complete script to extract everything in one go:

#!/bin/bash
# extract_all_models.sh — Pull all AI models from Insta360 Wave
set -e

OUTDIR="${1:-insta360_wave_models}"
mkdir -p "$OUTDIR"

echo "=== Step 1: Enable ADB ==="
python3 -c "
import hid
h = hid.device()
h.open(0x18F0, 0x4E40)
h.set_nonblocking(1)
while h.read(64): pass
pkt = bytearray(64)
pkt[0]=0x03; pkt[1]=0xF0; pkt[2]=0x00; pkt[4]=0xF0; pkt[5]=0xC0
h.write(bytes(pkt))
h.close()
print('ADB enabled')
"
sleep 5
adb wait-for-device

echo "=== Step 2: Pull ASR models (4.1 GB, ~10 min) ==="
adb pull /model/ "$OUTDIR/model/"

echo "=== Step 3: Pull TEA SDK models (75 MB) ==="
adb pull /usr/bin/tea_models/ "$OUTDIR/tea_models/"

echo "=== Step 4: Pull S2T data (1 MB) ==="
adb pull /usr/lib/s2t_data/ "$OUTDIR/s2t_data/"

echo "=== Step 5: Pull runtime libraries ==="
mkdir -p "$OUTDIR/runtime"
for lib in libonnxruntime.so.1.15.1 libnnsdk.so libasr_psp.so libasr_reverb.so libTEAProcessSdk.so; do
    adb pull "/usr/lib/$lib" "$OUTDIR/runtime/" 2>/dev/null || true
done

echo "=== Step 6: Restore normal USB ==="
python3 -c "
import hid
h = hid.device()
try:
    h.open(0x18F0, 0x4E40)
    h.set_nonblocking(1)
    while h.read(64): pass
    pkt = bytearray(64)
    pkt[0]=0x03; pkt[1]=0xF0; pkt[2]=0x00; pkt[4]=0xF0; pkt[5]=0xE0
    h.write(bytes(pkt))
    h.close()
    print('Normal USB restored')
except:
    print('Device already in normal mode')
"

echo "=== Done! ==="
echo "Models extracted to: $OUTDIR"
du -sh "$OUTDIR"
find "$OUTDIR" -type f | wc -l
echo "files total"

chmod +x extract_all_models.sh
./extract_all_models.sh my_wave_models

---

Notes

• ADB mode is not persistent — it resets on device reboot. Re-run the enable script after each reboot.
• USB 2.0 speed — The Wave connects at High Speed (480 Mbps), so 4 GB takes ~10 minutes.
• No impact on device — Pulling files is read-only. The device continues operating normally.
• Model partition is read-only — Mounted as ext4 RO. You cannot modify models via ADB without remounting.
• .nb files are hardware-specific — They only run on the Amlogic A5 NPU. The .onnx files are portable.