8. EdgeAI Model Zoo

The EdgeAI ModelZoo delivers out-of-box audio deep-learning demos that run directly on the AM62Dx target. Each reference model ships as an ONNX graph plus, where applicable, a pre-compiled TVM-RT + TIDL artifact, so you can reproduce these examples on the AM62D without compiling anything yourself. Two demo front-ends are provided: standalone Python inference scripts and Jupyter notebooks. The same on-target runtime hosts your own compiled artifacts once you build them through the x86 flow. This section runs entirely on the target; the x86 compile-and-evaluate flow is covered in Model Compile and Evaluate.

Note

The models in this repository are made available for experimentation and development. They are not intended for deployment in production.

8.1. What ships

Model	Task	Model ID	On target	Demo
VGGish11	Sound classification (UrbanSound8K)	`ac-0101`	C7™ NPU (TVM-RT + TIDL, 8-bit)	Script + notebook
YAMNet	Sound classification (UrbanSound8K)	`ac-0201`	C7™ NPU (TVM-RT + TIDL, 8-bit)	Script + notebook
GCRN	Speech enhancement (VoiceBank-DEMAND-16k)	`ase-0201`	C7™ NPU (TVM-RT + TIDL, 16-bit)	Script + notebook

VGGish11, YAMNet, and GCRN each ship a pre-compiled TVM-RT + TIDL artifact and run offloaded to the C7™ NPU.

8.2. On-target setup

The steps below run on the AM62D target (aarch64), on the Linux command line of the Processor SDK rootfs. The notebooks and scripts run in a Python virtual environment that reuses the TIDL-enabled TVM runtime already present in the rootfs.

Clone the repository. Clone the AudioAI ModelZoo repository (https://github.com/TexasInstruments/audioai-modelzoo):

$ mkdir -p ~/tidl && cd ~/tidl
$ git clone https://github.com/TexasInstruments/audioai-modelzoo.git
$ cd audioai-modelzoo

Download the model artifacts. Downloads the three pre-compiled TVM-RT + TIDL artifacts (VGGish11, YAMNet, GCRN)
```
$ ./download_models.sh -y
```
The script defaults to an interactive selection menu; the -y flag downloads everything non-interactively. Add -l to print the source URLs without downloading.
Create the virtual environment. Creates the venv at ~/venv/modelzoo with --system-site-packages so the TIDL-enabled runtime from the rootfs is visible:
```
$ ./setup_venv.sh
```
Activate the venv. Activate it in every new shell before running the demos:
```
$ source ~/venv/modelzoo/bin/activate
```

8.3. Inference script demos

Run each script from its own model directory in the activated venv. All four models ship a command-line script.

8.3.1. YAMNet (sound classification)

$ cd ~/tidl/audioai-modelzoo/inference/yamnet_sc
$ python3 yamnet_infer_audio.py --audio-file samples/miaow_16k.wav

The script converts the audio into log-mel patches, runs each patch through the TVM-RT + TIDL session, and averages the per-patch scores. For the bundled miaow_16k.wav clip the top classes are cat/animal sounds. Expected output (abridged):

======================================================================
YAMNET INFERENCE SUMMARY
======================================================================

Audio File: miaow_16k.wav
Artifacts: ac-0201_tvmrt_audio_classification_urbansound8k_yamnet_onnx
Audio Duration: 6.73 seconds
Number of Patches: 7

======================================================================
TOP 10 PREDICTIONS (averaged across all patches):
======================================================================
Rank   Score      Class
----------------------------------------------------------------------
1      ...        Cat
2      ...        Animal
...
======================================================================
PERFORMANCE METRICS:
======================================================================
Inference Time (all patches): 17.53 ms total | 7 patches
RTF = 0.0026 (0.018s / 6.73s)
======================================================================

A second sample, samples/speech_whistling2.wav, is also included.

8.3.2. VGGish11 (sound classification)

$ cd ~/tidl/audioai-modelzoo/inference/vggish11_sc
$ python3 vggish_infer_audio.py --audio-file sample_wav/139951-9-0-9.wav

The bundled clip is UrbanSound8K class 9 (street music). Expected output (abridged):

======================================================================
VGGISH11 INFERENCE SUMMARY
======================================================================

Audio File: 139951-9-0-9.wav
Artifacts: ac-0101_tvmrt_audio_classification_urbansound8k_vggish11_onnx
Audio Duration: 4.00 seconds
Top Class: Street music (Class 9) - Score: ...

Performance:
  Inference: 8.88 ms
  Throughput: RTF = 0.0022 (0.009s / 4.00s)
======================================================================

8.3.3. GCRN (speech enhancement)

$ cd ~/tidl/audioai-modelzoo/inference/gcrn_se
$ python3 gcrn_infer_audio.py --input sample_wav/noisy.wav

The script writes the enhanced audio to enhanced.wav beside the input and a per-block timing report to gcrn-benchmark.md. Pass --output to change the destination wav. Expected output (abridged):

Machine: aarch64  SOC: am62d  TIDL_VER: ...
Loading TVM + TIDL session...
Running inference on noisy.wav (9.77s)...
Wrote enhanced audio to sample_wav/enhanced.wav

GCRN TVM+TIDL benchmark (3 chunk(s), 9.77s audio)

block                    total (ms)  per-chunk (ms)
...
Real-Time Factor (inference / audio) = 3.9500

Wrote benchmark report to sample_wav/gcrn-benchmark.md

8.4. Jupyter notebook demos

All four reference models have a Jupyter notebook demo. Start Jupyter Lab from the activated venv:

$ cd ~/tidl/audioai-modelzoo
$ ./jupyter_lab_venv.sh

The launcher prints a highlighted access URL. Open the URL in a browser on a machine that can reach the target and enter the token tidl when prompted:

Access URL: http://<target-ip>:8888/lab?token=tidl

Jupyter Lab opens with the GCRN notebook loaded. To run a different model, pick its notebook from the file browser in the left sidebar. The notebooks are:

Model	Notebook
GCRN	`inference/gcrn_se/gcrn_inference.ipynb`
VGGish11	`inference/vggish11_sc/vggish_inference.ipynb`
YAMNet	`inference/yamnet_sc/yamnet_inference.ipynb`

8.5. Compiled artifacts

download_models.sh places the artifacts under model_artifacts/<TIDL_VER>/am62d/. Each model gets its own directory, named after its model ID, containing a model/ folder (the trained ONNX model) and, where applicable, an artifacts/ folder (the compiled model artifacts). VGGish11, YAMNet, and GCRN each get a pre-compiled TVM-RT + TIDL artifact:

ac-0101 - ac-0101_tvmrt_audio_classification_urbansound8k_vggish11_onnx (VGGish11)
ac-0201 - ac-0201_tvmrt_audio_classification_urbansound8k_yamnet_onnx (YAMNet)
ase-0201 - ase-0201_tvmrt_speech_enhancement_voicebank_demand_16k_gcrn_fixed_4sec_onnx (GCRN)

Each of those artifact directories’ artifacts/ folder holds the TVM-RT deployment artifacts (deploy_lib.so, deploy_graph.json, deploy_param.params) in both target (.evm, aarch64) and host (.pc, x86) variants. The demo scripts and notebooks select the .evm variant automatically on the target.