Element 22.10
Microphones, acoustic sensing and local speech
- Sheet
- 22.10
- State
- Drafted
- Release
- 16 November 2026
- Editor
- Contributed by Praecise as open research
- Reviewers
- To be named at review
- Safety
- S1, as the chapter
- Licence
- CC BY-SA 4.0
- Resources
- 12
The element
- Requirement
Sound is recorded only for a stated purpose: wildlife, machine health, voice control or accessibility. Recognition runs locally where it can, and recording of people's speech in shared spaces is avoided or governed.
- How to tell
- The recorder's sample rate fits the sound. An open logger records 8,000 to 384,000 samples per second, reaching ultrasonic frequencies, uncompressed to microSD (AudioMoth).
- Speech recognition runs without the internet where privacy matters. Offline toolkits state models of about 50 MB per language for more than 20 languages (Vosk), or memory from about 273 MB to 3.9 GB depending on model size (whisper.cpp).
- The model's training and limits are known. Whisper was trained on 680,000 hours of multilingual audio; its authors report accuracy approaching human levels on benchmarks (Robust Speech Recognition via Large-Scale Weak Supervision (Whisper)). Accuracy in local accents and languages is tested, not assumed.
- For machine monitoring, a detector is trained on normal sound and scored by how well it catches faults at low false-alarm rates, as the DCASE challenge does (DCASE 2024 Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring).
- Options
- Passive acoustic monitoring for wildlife: open loggers (AudioMoth) with bird sound identification software (BirdNET-Analyzer).
- Microphone arrays to find where a sound comes from and separate sources: an open library with open 8- and 16-input array designs (ODAS (Open embeddeD Audition System)).
- Local speech-to-text on a single-board computer or phone (whisper.cpp, Vosk).
- Keyword spotting on microcontrollers: an open dataset built for it (Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition) and an inference framework for devices with tens to hundreds of kilobytes of memory (TensorFlow Lite for Microcontrollers); a benchmark suite compares accuracy, latency and energy on such devices (MLPerf Tiny Benchmark (paper)).
- Machine sound monitoring: open datasets of normal and faulty valves, pumps, fans and slide rails (MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection) and the DCASE anomaly task (DCASE 2024 Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring).
- MEMS microphone specifications: no maker datasheet could be opened in this work; see the research note.
- Settled, open, contested
Settled: offline speech recognition on small computers is practical (whisper.cpp, Vosk). Settled: anomaly detection trained only on normal sounds is an active, benchmarked field (DCASE 2024 Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring). Open: accuracy of local speech recognition for languages and accents with little training data. Contested: always-listening devices. Their convenience is real; so is the risk that conversations are captured. No source specific to audio recording law was opened in this work; the European principles for video devices on signage, necessity and short retention (Guidelines 3/2019 on processing of personal data through video devices) are the nearest reference. Read local law.
The four stages
- Learn: Robust Speech Recognition via Large-Scale Weak Supervision (Whisper); DCASE 2024 Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring.
- Design: choose sample rate, placement and what is kept; for keyword spotting, Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition and MLPerf Tiny Benchmark (paper).
- Build: AudioMoth, ODAS (Open embeddeD Audition System), TensorFlow Lite for Microcontrollers.
- Deploy: BirdNET-Analyzer, whisper.cpp, Vosk, with retention and deletion rules.
Resources at each stage
1 Learn
Understand
- DCASE 2024 Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring Dataset, S0
- MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection Dataset, S0
- Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition Dataset, S0
- Robust Speech Recognition via Large-Scale Weak Supervision (Whisper) Paper, S0
2 Design
Size and choose
- DCASE 2024 Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring Dataset, S0
- Guidelines 3/2019 on processing of personal data through video devices Guide, S1
- MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection Dataset, S0
- MLPerf Tiny Benchmark (paper) Paper, S0
- Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition Dataset, S0
- TensorFlow Lite for Microcontrollers Library, S0
3 Build
Build or obtain
- AudioMoth Reference design, S0
- BirdNET-Analyzer Tool, S0
- ODAS (Open embeddeD Audition System) Library, S0
- TensorFlow Lite for Microcontrollers Library, S0
- Vosk Tool, S0
- whisper.cpp Tool, S0
4 Deploy
Install, run and measure
- AudioMoth Reference design, S0
- BirdNET-Analyzer Tool, S0
- Guidelines 3/2019 on processing of personal data through video devices Guide, S1
- Vosk Tool, S0
- whisper.cpp Tool, S0
What changes between 30 adults and 300 people
At 30 adults, a few loggers on the land and voice control on one machine are enough. At 300 people, a monitoring programme across habitats, acoustic monitoring of pumps and generators, and voice interfaces for people who cannot use screens become worthwhile, with a written rule that speech is not recorded in shared spaces.
Reference designs
-
Reference design
Low-cost open acoustic logger from Open Acoustic Devices that records uncompressed audio to microSD at 8,000 to 384,000 samples per second, reaching ultrasonic frequencies, on three AA batteries, with waterproof (IPX7) and underwater cases.
External resource. Not made or verified by Public ProtocolMaturity: ReplicatedSafety S0Open, with paid products or servicesListed 16 November 2026
Parts list
Every resource that supports this element. Those marked External carry the label “External resource. Not made or verified by Public Protocol”.
| Resource | Type | Origin | Maturity | Safety | Licence, as stated at its source |
|---|---|---|---|---|---|
| AudioMoth | Reference design | External | Replicated | S0 | Not stated on the page opened |
| BirdNET-Analyzer | Tool | External | Established | S0 | Source code: MIT License. Models: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License |
| DCASE 2024 Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring | Dataset | External | Working | S0 | Not stated on the page opened |
| Guidelines 3/2019 on processing of personal data through video devices | Guide | External | Established | S1 | Not stated on the page opened |
| MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection | Dataset | External | Established | S0 | Not stated on the page opened |
| MLPerf Tiny Benchmark (paper) | Paper | External | Established | S0 | Not stated on the page opened |
| ODAS (Open embeddeD Audition System) | Library | External | Working | S0 | MIT |
| Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition | Dataset | External | Established | S0 | CC BY 4.0 |
| TensorFlow Lite for Microcontrollers | Library | External | Established | S0 | Apache-2.0 |
| Vosk | Tool | External | Established | S0 | Apache-2.0 |
| whisper.cpp | Tool | External | Established | S0 | MIT |
| Robust Speech Recognition via Large-Scale Weak Supervision (Whisper) | Paper | External | Established | S0 | arXiv non-exclusive distribution licence 1.0 |
Depends on, and completes
It is one of 12 elements that complete Delivery, operations and measurement.
The standing questions
Who hears what is recorded. Who decides which spaces are free of microphones. Who checks that a voice system understands every member.
Environments
Wet and underwater
a waterproof IPX7 case with an acoustic vent, and an underwater case tested to 30 m for two months (AudioMoth).
Tropical humid forest, temperate forest
wildlife monitoring is the documented use (AudioMoth, BirdNET-Analyzer); no source found on humidity damage.
- Islands and coasts, arctic cold, hot deserts, high altitude, flood plains: no source found.
Environment profiles for these settings: Tropical rainforest and humid tropics; Temperate and boreal forests; High humidity. All environments
Safety
State and evidence
Drafted
An editor has written the element's text from the resources listed under it, citing them. It is checked when it is reviewed.
Evidence from a real place or system is recorded here, with what did not work. What the states mean
History
- Drafted in the release of 16 November 2026.