MultiNet Alternative for ESP32: One TFLite Model, Multiple Command Words

ESP-SR MultiNet is an offline command-word architecture. Voicute provides the same useful device pattern with a different delivery path: one compact INT8 TFLite model recognizes multiple fixed keywords and returns a command ID for ESP32 firmware.

The seven supported languages

The current Voicute language matrix is Chinese, English, Japanese, French, German, Spanish and Hindi. The same Multi-Keyword workflow can create a model for one selected language and several fixed commands.

How one model replaces a group of command detectors

Voicute Multi-Keyword is not a bundle of independent single-keyword models. Keywords share feature extraction and one inference pass distinguishes the configured set. The model output maps each recognized keyword to a command ID, similar to the command-word role played by MultiNet.

ESP32 deployment with TFLite Micro

Export the matched INT8 TFLite model and output head, keep the 16 kHz mono feature pipeline unchanged, and integrate both into an ESP-IDF firmware. A wake detector can start the interaction; the Multi-Keyword model then recognizes the fixed command set.

Language coverage beyond ESP-SR

If ESP-SR MultiNet does not provide the language model you need, a language-specific TFLite model avoids forcing translated text into the wrong acoustic model. The same approach applies to the seven Voicute languages; additional languages require their own training and hardware evaluation.

Build the demo

Start with the open-source ESP32 example and the public multi-keyword demo model. Then test the final microphone, enclosure, noise conditions, recall, false triggers, latency and RAM before shipping.

ESP32 demo deployment

git clone https://github.com/voicute/onnx-wakeword.git\ncd onnx-wakeword/esp32/examples/esp32s3_hmi_devkit\nidf.py set-target esp32s3\nidf.py build

Start with the open-source ESP32 example and the public multi-keyword demo model. Validate the final microphone, enclosure, noise conditions, recall, false triggers, latency and RAM before shipping.