| --- |
| license: apache-2.0 |
| library_name: nemo |
| tags: |
| - onnx |
| - nemo |
| - speech-commands |
| - wake-word-spotting |
| datasets: |
| - HashNuke/tincan-wakewords-data |
| metrics: |
| - accuracy |
| model-index: |
| - name: TinCan Speech Commands Model |
| results: |
| - task: |
| type: audio-classification |
| name: Speech command recognition |
| dataset: |
| name: TinCan Speech Commands validation set |
| type: tincan-speech-commands-validation |
| metrics: |
| - type: loss |
| name: Validation loss |
| value: 0.1493 |
| - type: accuracy |
| name: Validation micro top-1 accuracy |
| value: 95.28 |
| - type: accuracy |
| name: Validation macro accuracy |
| value: 94.61 |
| --- |
| |
| # TinCan Speech Commands Model |
|
|
| A compact English speech-command recognition model for tincan app. |
|
|
| This model recognizes 47 short command classes and is designed for small-footprint command recognition where cloud ASR is unnecessary or undesirable. The exported ONNX artifact is under 400 KB, making it practical for local-first applications, prototypes, and edge deployments. |
|
|
| * 12 custom words |
| * and 35 words from the Google Speech Commands dataset v2 |
|
|
| ## Highlights |
|
|
| - 47-class English command recognizer |
| - ONNX export for portable inference |
| - Small model artifact: `model.onnx` is approximately 378 KB |
| - Based on NVIDIA NeMo's MatchboxNet command-recognition model family |
|
|
| ## Base Model |
|
|
| This model uses NVIDIA NeMo's `commandrecognition_en_matchboxnet3x2x64_v2` MatchboxNet command-recognition architecture. |
|
|
| Base model reference: [`commandrecognition_en_matchboxnet3x2x64_v2`](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/commandrecognition_en_matchboxnet3x2x64_v2) |
|
|
| ## Metrics |
|
|
| These metrics describe the currently exported `model.onnx` artifact. |
|
|
| | Metric | Value | |
| |---|---:| |
| | Validation loss | 0.1493 | |
| | Validation micro top-1 accuracy | 95.28% | |
| | Validation macro accuracy | 94.61% | |
|
|
| ## Supported Commands |
|
|
| Custom TinCan commands: |
|
|
| `astra`, `bali`, `boston`, `capri`, `delhi`, `dublin`, `frisco`, `monaco`, `oslo`, `paris`, `seatown`, `tokyo` |
|
|
| Google Speech Commands labels: |
|
|
| `yes`, `no`, `up`, `down`, `left`, `right`, `on`, `off`, `stop`, `go`, `zero`, `one`, `two`, `three`, `four`, `five`, `six`, `seven`, `eight`, `nine`, `bed`, `bird`, `cat`, `dog`, `happy`, `house`, `marvin`, `sheila`, `tree`, `wow`, `backward`, `forward`, `follow`, `learn`, `visual` |
|
|
| ## Inference Notes |
|
|
| The model outputs logits over the 47 labels listed in `labels.json`. Use the output index to look up the predicted command label. |
|
|
| ## Training Provenance |
|
|
| | Field | Value | |
| |---|---| |
| | Model name | `commandrecognition_en_matchboxnet3x2x64_v2` | |
| | Export format | ONNX | |
| | Epochs | 10 | |
| | Batch size | 32 | |
|
|
| ## Limitations |
|
|
| - This is a closed-vocabulary command recognizer, not a general speech-to-text model. |
| - The model is intended for English short-command recognition. |
| - Validation metrics may not fully predict performance with every microphone, speaker, accent, room, or noise condition. |
|
|
