Precise “Hey Bobo” Wake-Word
Budget: ₹12,500 – ₹37,500 INR
I need a production-ready wake-word model that reliably responds to “Hey Bobo” on both ESP32 and Raspberry Pi boards. Accuracy is critical: in a moderately noisy room it should trigger every time someone clearly says the phrase, yet ignore most unrelated speech. I still want slight phonetic wiggle-room—utterances such as “hello bo,” “hey bebo,” or “hello bebo” should wake the system as well—so fine-tuned threshold settings and a well-balanced dataset will be essential.
What I expect from you
• A trained wake-word model or firmware that runs in real time on ESP-IDF for ESP32 and on Raspbian/Ubuntu for Raspberry Pi without requiring cloud calls.
• Demo code that shows how to load the model, stream audio from an onboard mic, and raise an event when the wake word (or approved variants) is detected.
• Instructions for collecting additional audio samples so I can keep refining the model if needed, plus clear guidelines on adjusting sensitivity to maintain high precision while preventing false positives.
• CPU- and memory-usage figures on each platform so I know exactly how lightweight the solution is.
Acceptance criteria
1. Latency from spoken phrase to callback ≤ 250 ms on both targets.
2. ≥ 95 % wake-up rate in moderate background noise (office chatter, music at low volume).
3. ≤ 2 false activations per hour of continuous speech.
4. Complete build steps and source so I can reproduce the binary from scratch.
You are free to use tools like TensorFlow Lite Micro, Porcupine, Picovoice DIY, or a custom DSP pipeline—as long as licensing permits commercial use. I am happy to provide extra voice recordings to help you fine-tune. If you’ve shipped similar models before, let me know; a quick video demo on actual hardware will fast-track selection.
data collection and everything should be done solely by the freelanacer's job.
What I expect from you
• A trained wake-word model or firmware that runs in real time on ESP-IDF for ESP32 and on Raspbian/Ubuntu for Raspberry Pi without requiring cloud calls.
• Demo code that shows how to load the model, stream audio from an onboard mic, and raise an event when the wake word (or approved variants) is detected.
• Instructions for collecting additional audio samples so I can keep refining the model if needed, plus clear guidelines on adjusting sensitivity to maintain high precision while preventing false positives.
• CPU- and memory-usage figures on each platform so I know exactly how lightweight the solution is.
Acceptance criteria
1. Latency from spoken phrase to callback ≤ 250 ms on both targets.
2. ≥ 95 % wake-up rate in moderate background noise (office chatter, music at low volume).
3. ≤ 2 false activations per hour of continuous speech.
4. Complete build steps and source so I can reproduce the binary from scratch.
You are free to use tools like TensorFlow Lite Micro, Porcupine, Picovoice DIY, or a custom DSP pipeline—as long as licensing permits commercial use. I am happy to provide extra voice recordings to help you fine-tune. If you’ve shipped similar models before, let me know; a quick video demo on actual hardware will fast-track selection.
data collection and everything should be done solely by the freelanacer's job.