| --- |
| title: Tiny Hinglish Turn Detector |
| emoji: 🎙️ |
| colorFrom: yellow |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 5.49.1 |
| python_version: "3.12" |
| app_file: app.py |
| pinned: false |
| license: other |
| short_description: Tiny HOLD/END detection for Indian Hinglish voice agents |
| --- |
| |
| # Tiny Hinglish Turn Detector |
|
|
| > **Development preview:** one of 83 upstream training shards was available; |
| > no official-test or collected-Hinglish performance is claimed. The packaged |
| > 151,812-parameter TinyTCN is fast, but the acoustic logistic baseline is |
| > stronger on the current IID development split. |
|
|
| This Space runs the packaged ONNX model at one VAD pause checkpoint per click |
| and uses the same serialized default decision configuration as the replay tool. |
| Stateful latching and multi-checkpoint behavior are validated through replay; |
| the UI itself is a single-checkpoint visualizer. |
|
|
| The model uses a four-second suffix and a stored threshold of `0.7410007715`. |
| On 326 adaptively reused development examples it reaches AP 0.6881 and AUROC |
| 0.7390, with 10 TP / 3 FP / 163 TN / 150 FN. Treat the output as an inspectable |
| demo, not a production endpointing claim. |
|
|
| Repository-authored source code is Apache-2.0. The preview weights carry no |
| license grant while upstream-derived-weight rights remain unresolved. Upstream |
| training audio is not redistributed. |
|
|
