Keyword Spotting and Audio Event Detection on Microcontrollers
Always-on speech command, wake-word, noise suppression and sound event models small enough for microcontrollers and DSPs.
Tier Grades
Tier A: Frontier AI labs
feasible
Arm, Google and Qualcomm publish the reference keyword spotters (Hello Edge, TFLite Micro).
Tier B: Indonesian AI labs
feasible
Training a DS-CNN keyword spotter is small; Indonesian keyword data is not in the sources.
Tier C: Institution units such as ITB
feasible
Hello Edge came from a 4-author team.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Vendor keyword spotters run on low-cost boards for voice-control proofs of concept.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeHello Edge: architectures fitted to microcontroller memory and compute; DS-CNN reaches 95.4% accuracy, about 10% above a DNN with a similar parameter count.
- DataSpeech command datasets for keyword spotting (Hello Edge, NXP and Silicon Labs model zoos); the dataset size is not in the abstract.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 4 authors on Hello Edge.
- Time horizonOne to two years.
- Shelf-life signalsKeywords in local languages, which vendor zoos do not cover (vendor examples use English words), Audio event detection and noise suppression joining wake words on the same chips
- Pipeline roleTier A and vendors publish reference models; Tier C collects Indonesian keyword audio; Tier D builds voice-controlled devices in proofs of value.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
DS-CNNTinyMLexemplarHello Edge keyword spotting modelsTinyMLexemplar
MLPerf Tiny DS-CNN (KWS)TinyMLexemplar
Micro Speech (TFLM keyword model)TinyMLexemplar
BC-ResNet (Broadcasted Residual Network)TinyMLproduced byTENet (Temporal Efficient Neural Network)TinyMLproduced byHello Edge KWS CNN (small, medium, large)TinyMLproduced byHello Edge KWS DNN (small, medium, large)TinyMLproduced byMicroNetTinyMLproduced byMLPerf Tiny 1D DS-CNN (streaming wakeword)TinyMLproduced by
Microspeech LSTMTinyMLproduced by
Keyword Spotting NumbersTinyMLproduced byKeyword Spotting On/Off v3TinyMLproduced bySilicon Labs MLTK keyword spotting modelsTinyMLproduced by
ST MiniResNet v1 and v2 audio event detectionTinyMLproduced byWav2letterTinyMLproduced byLinks
R&D TopicKeyword Spotting and Audio Event Detection on Microcontrollers
ModelsDS-CNNArm, TinyMLMLPerf Tiny DS-CNN (KWS)MLCommons, TinyMLHello Edge keyword spotting modelsArm, TinyMLMicro Speech (TFLM keyword model)Google, TinyML+12 more
DS-CNNArm, TinyMLMLPerf Tiny DS-CNN (KWS)MLCommons, TinyMLHello Edge keyword spotting modelsArm, TinyMLMicro Speech (TFLM keyword model)Google, TinyML+12 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 16
Hello Edge keyword spotting modelsArm, TinyMLexemplarSilicon Labs MLTK keyword spotting modelsSilicon Labs, TinyMLproduced byHello Edge KWS CNN (small, medium, large)Arm, TinyMLproduced byHello Edge KWS DNN (small, medium, large)Arm, TinyMLproduced byKeyword Spotting NumbersSilicon Labs, TinyMLproduced byKeyword Spotting On/Off v3Silicon Labs, TinyMLproduced byBC-ResNet (Broadcasted Residual Network)Qualcomm, TinyMLproduced byDS-CNNArm, TinyMLexemplarMicro Speech (TFLM keyword model)Google, TinyMLexemplarMicrospeech LSTMNXP Semiconductors, TinyMLproduced byMLPerf Tiny 1D DS-CNN (streaming wakeword)MLCommons, TinyMLproduced byMLPerf Tiny DS-CNN (KWS)MLCommons, TinyMLexemplarPapers
- Hello Edge: Keyword Spotting on MicrocontrollersYundong Zhang et al., arXiv preprint, 2017 founding
- Speech Commands: A Dataset for Limited-Vocabulary Speech RecognitionPete Warden, arXiv preprint, 2018 founding
- Keyword Transformer: A Self-Attention Model for Keyword SpottingAxel Berg et al., Proc. Interspeech 2021, 4249-4253, 2021 state of the art
- Broadcasted Residual Learning for Efficient Keyword SpottingByeonggeun Kim et al., INTERSPEECH 2021 (per arXiv comment), 2021 state of the art
- MatchboxNet: 1D Time-Channel Separable Convolutional Neural Network Architecture for Speech Commands RecognitionSomshubra Majumdar et al., arXiv preprint, 2020 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/1711.07128 https://arxiv.org/abs/1711.07128
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Keyword Spotting and Audio Event Detection on Microcontrollers", https://aurion.semai.id/topics/keyword-spotting-and-audio-on-microcontrollers/, records of October 2026.