🧪 Real-Time Demo (Speaker Verification - ID)

This page explains how to run a real-time Speaker Verification (ID) demo using a trained model. The demo supports both embedded (EVB) and PC-based platforms and demonstrates live speaker enrollment and identity matching.

🔧 Run `demo` Mode

soundkit -t id -m demo -c configs/id/id.yaml demo.platform=pc # or evb

🧾 Demo Parameters

Parameter	Description
`platform`	`pc` or `evb`: run the model on desktop or Ambiq hardware
`epoch_loaded`	Model checkpoint to use for inference (`best`, `latest`, or a specific epoch number)
`tflite_dir`	Directory containing the exported `.tflite` model
`evb_dir`	Path to the EVB firmware project (used for build and deploy)
`pre_gain`	Optional gain multiplier applied before inference
`filename`	Output filename used for generated C header or binary
`param_struct_name`	C structure name for exported model parameters
`num_utterances_registered`	Number of utterances used per speaker for enrollment
`frames_vad_trigger_id`	Frame count to trigger ID matching after VAD confirms speech presence

Example:

demo:
  platform: pc
  epoch_loaded: best
  tflite_dir: ./soundkit/tasks/id/tflite
  evb_dir: ./soundkit/tasks/id/evb
  pre_gain: 1
  filename: def_nn2_nnid
  param_struct_name: params_def_nn2_nnid
  num_utterances_registered: 4
  frames_vad_trigger_id: 180

💻 Deployment Modes

Example

PC

Run on the PC-based test for real-time ID task.

soundkit -t id -m demo -c configs/id/id.yaml demo.platform=pc # run on PC

EVB

Run on the PC-based test or Ambiq's family of ultra-low power SoCs for real-time KWS task. You can choose on PC or EVB based on command-line overwrite

Type

soundkit -t id -m demo -c configs/id/id.yaml demo.platform=evb # run on Ambiq EVB

Go to https://ambiqai.github.io/soundkit/web-dashboards/id-usb/
Follow the GUI instructions
If you have any connection issue or no waveform showing. See the Troubleshooting

✅ Output

Real-time verification results (e.g., match/no match)
Optional logging for scores, decision thresholds, and debug stats
Timing benchmarks and event indicators for latency analysis

🧠 Tips

Use pre_gain to adjust microphone input levels for better accuracy
Ensure tflite_dir and epoch_loaded match the latest export
Re-run export if model architecture or parameters have changed
Adjust num_utterances_registered to fine-tune enrollment stability