“Voice Interfaces on a Budget: Building Real-time Speech Recognition on Low-cost Hardware,” a Presentation from Useful Sensors
For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2025/07/voice-interfaces-on-a-budget-building-real-time-speech-recognition-on-low-cost-hardware-a-presentation-from-useful-sensors/ Pete Warden, CEO of Useful Sensors, presents the “Voice Interfaces on a Budget: Building Real-time Speech Recognition on Low-cost Hardware” tutorial at the May 2025 Embedded Vision Summit. In this talk, Warden presents Moonshine, a speech-to-text model that outperforms OpenAI’s Whisper by a factor of five in terms of speed. Leveraging this efficiency, he shows how to build a voice interface on a low-cost, resource-constrained Cortex-A SoC using open-source tools. He also covers how to use voice activity detection as a first step before running speech-to-text to avoid false positives on noise that isn’t speech. In addition, he demonstrates how to use Python to control speech recognition and take actions based on recognized words. The Moonshine model’s compact size (as small as 26 MB) and high accuracy (<5% word error rate) make it ideal for embedded applications. All code and documentation are available online, allowing you to replicate the project. This presentation showcases the potential for voice-enabled interfaces on affordable hardware, enabling a wide range of innovative applications.
“Voice Interfaces on a Budget: Building Real-time Speech Recognition on Low-cost Hardware,” a Presentation from Useful Sensors
- 1. Voice Interfaces on a Budget: Building Real-Time Speech Recognition on Low-Cost Hardware Pete Warden CEO Useful Sensors
- 2. What Is This Talk About? • Voice interfaces in the past: • Cost $$$ • Only available to big tech co’s • Required specialists • Took years to build • Voice interfaces now: • Open source • Available to everyone • Usable by any software engineer © 2025 Useful Sensors Inc 2
- 3. What You’ll Leave With • How to build a simple voice app • Running on low-cost hardware (Raspberry Pi) • Without a cloud API or network connectivity • Using open-source, freely available software and models © 2025 Useful Sensors Inc 3
- 4. Does Anyone Use Voice Interfaces? • Siri, Alexa, Ok Google? • No, except for timers and baby shark • However, a thought experiment: • You’re next to your significant other on a couch • Do you text them to decide what show to watch? • So, we do like voice interfaces, but the current ones aren’t good enough © 2025 Useful Sensors Inc 4
- 5. What Do You Need to Start? • Cortex A CPU or equivalent • MCUs soon, hopefully • Open source frameworks • Open weights speech recognition models © 2025 Useful Sensors Inc 5
- 6. Speech Recognition Models • OpenAI’s Whisper • First production-quality open weights ASR model • Smallest version is 40 million parameters • Can run on APUs, but hard to get real-time • Always processes 30 seconds of audio at once, very wasteful for interactive use cases • Useful Sensors’ Moonshine • Open weights • Achieves same accuracy as Whisper for tiny and base models • Smallest version is 26 million parameters • Able to run well on most modern APUs • Flexible input window, so you only compute what you need © 2025 Useful Sensors Inc 6
- 7. Moonshine Options • Tiny version: • 26 million parameters • Word error rate of 4.51% • Base version: • 52 million parameters • Word error rate of 3.29% • Tutorial uses Tiny • Many frameworks supported: • PyTorch • Keras • TensorFlow • ONNX • We’re using ONNX • Quantized versions available: • 26 MB / 56 MB file sizes • 1.6x faster than float © 2025 Useful Sensors Inc 7
- 8. Now I’ll show you how to run an interactive speech application using Moonshine on a Raspberry Pi 5. Based on material from my Stanford EE292D Edge AI course: https://github.com/ee292d/labs/tree /main/lab4 You’ll need a Pi 5, some way to connect to it, and a USB microphone. © 2025 Useful Sensors Inc Tutorial 8
- 9. Live Coding Demo © 2025 Useful Sensors Inc 9
- 10. Next Steps • How can you take action based on speech? • Plain old string matching can work for simple uses • Recognizing a natural speaking style needs speech to intent • Still a research problem • https://github.com/AIWintermuteAI /Speech-to-Intent-Micro • What about text to speech? • Speakers are a cheaper alternative to displays • PiperTTS is very efficient, runs on Pi’s, and sounds good • Hyper-realistic models are emerging, but they use a lot of resources, won’t work on a Pi (yet) © 2025 Useful Sensors Inc 10
- 11. Conclusion • It’s never been easier to build a voice-driven product • It’s still early days for voice, don’t write it off because Siri isn’t popular • You’ve got this! © 2025 Useful Sensors Inc 11
- 12. Whisper: https://openai.com/index/whisper/ Moonshine: https://github.com/usefulsensors/mo onshine PiperTTS: https://piper.ttstool.com/ Speech-to-intent Micro: https://github.com/AIWintermuteAI/S peech-to-Intent-Micro EE292D Tutorial: https://github.com/ee292d/labs/tree/ main/lab4 Me: [email protected] © 2025 Useful Sensors Inc Resources 12