What Lyric moves before digitization
A conventional speech front end captures a 16-kHz waveform, then computes a 40-bin log-Mel spectrogram using a 50-ms window and a 10-ms hop, producing 100 frames per second. Before any classifier runs, each overlapping frame needs windowing, an FFT, power-spectrum computation, Mel filtering, logarithmic compression, and normalization. Shrinking the classifier does not remove that preprocessing.
Lyric instead performs frequency selection and envelope extraction in the acoustic and analog domain, before the analog-to-digital converters. Airborne speech drives five passive Helmholtz resonators, each tuned to a different region of the speech spectrum. A microphone sealed into each cavity measures its pressure response, and an analog AC-RMS circuit extracts the slowly varying amplitude over a 10-ms integration interval. The five envelopes are then digitized at 100 samples per second per channel, or 500 scalar samples per second in total, and passed to a causal temporal network. This path requires no digital framing, FFT, Mel filtering, or logarithmic compression.
Why channel identity replaces spectral bins
The design rests on a division of labor. Each resonator's channel identity encodes coarse frequency information, while its envelope captures how response amplitude changes over time. Because the ADCs sample only the slow envelopes, they never need to sample the acoustic carriers, which is why the 50-Hz Nyquist limit applies to envelope modulation even though the channel target frequencies extend to 7.5 kHz. Modulation above 50 Hz must still be attenuated by the analog envelope response before digitization.
The paper is explicit that only acoustic frequency selection is passive. The envelope signals still require appropriate filtering and sampling. The claim is therefore about a compact acoustic-to-digital representation and the digital processing it removes, not about eliminating analog circuitry.
How five channels were chosen
Channel selection had to balance redundancy against spectral coverage: similar channels carry overlapping information, while coverage gaps can erase distinctions between digits. The authors screened candidate banks in COMSOL simulation using a separate alphabet dataset before recording any AudioMNIST data. They simulated eight candidate resonators at 110, 180, 300, 1800, 3500, 4500, 6000, and 7500 Hz, computed peak-normalized transfer-energy responses, and predicted response trajectories from 16-kHz speech using a 40-ms Hann window, a 10-ms hop, and a 2048-point FFT.
All 56 five-resonator subsets were evaluated on 182 labeled A-Z segments from seven recordings, each subset represented by 25 features describing per-channel energy proportions and their temporal changes, using a class-balanced linear SVM with seven-fold leave-one-recording-out cross-validation. The subset targeting 300, 1800, 3500, 4500, and 7500 Hz ranked first at 67.03% mean balanced accuracy, against 66.48% for the runner-up, with a fold standard deviation of 15.23 percentage points. The authors note that the narrow margin and large fold variation do not establish a unique optimum, and that no digit data informed the choice.
The selected channels were built as monolithic PLA Helmholtz resonators, printing neck, cavity, transition, and microphone port as one part to eliminate assembly seams. Where prototype measurements were available, the measured peak was used to adjust cavity volume while holding the neck fixed; for the 1800-Hz channel, a predecessor peaking at 2150 Hz was corrected from 0.984 to 1.404 mL.
The accuracy and cost tradeoff
Evaluation used physically recorded AudioMNIST utterances with a speaker-disjoint split of 36 training, 12 validation, and 12 test speakers, across nine classifier families. Audio was peak-normalized to -9 dBFS and replayed 15 cm from the fixed resonator array, one speaker per session, with 150-ms gaps.
The headline paired comparison is EdgeSpeechNet-A. With Lyric input it reaches 95.70% accuracy at 6.56 ms and an estimated 1.97 mJ of dynamic processor energy per inference, against 99.28%, 454.31 ms, and 136.29 mJ for its log-Mel counterpart. That is roughly a 69-fold reduction in compute time and estimated energy for a 3.58-percentage-point accuracy loss. Across the seven families exceeding 95% with log-Mel, the five-channel input achieves 91.07-95.70% accuracy.
The two smallest classifiers move the other way. MicroCNN rises from 43.88% to 74.72% and StrideCNN from 47.52% to 80.92%, despite using fewer parameters. Replacing digital log-Mel extraction with envelope normalization cuts mean front-end time from 8.48 to 0.28 ms, and the five-channel input and its adapted networks require fewer MACs in every pair while retaining the 100-Hz input frame rate. Total compute latency falls by 69.5-98.6% across the nine families.
Where the latency gain is largest
The paired EdgeSpeechNet-A reduction is not the only comparison that matters. Five Lyric configurations exceed 90% accuracy in under 10 ms, and no evaluated log-Mel configuration meets both thresholds. Resonator PoolCNN reaches 91.07% at 1.81 ms and 0.54 mJ, while the cheapest log-Mel model above 90%, TC-ResNet8, requires 16.87 ms and 5.06 mJ.
Against that lower-cost baseline the picture changes. Resonator EdgeSpeechNet-A is 2.9 points below log-Mel TC-ResNet8, with a 2.6-fold reduction in compute time and estimated energy. The gain therefore persists against a cheaper log-Mel baseline, but it is smaller than the paired EdgeSpeechNet-A reduction. Energy estimates use the measured 0.30-W dynamic processor draw and exclude microphones, envelope circuitry, ADC, and acquisition time, so they describe processor computation rather than whole-system energy.
Why 100-Hz envelopes were enough
Two ablations test what the representation actually needs. Raising the resonator envelope rate from 100 to 400 Hz increases MACs by 3.8-4.0 times and total latency by 2.3-4.6 times, but changes mean accuracy across the seven stronger families only from 93.37% to 93.59%. MicroCNN and StrideCNN instead lose 7.45 and 4.57 points. The 100-Hz input avoids the extra compute cost without a consistent accuracy loss in these runs, though single-seed results cannot quantify run-to-run variation.
The second ablation addresses spectral selectivity. At 400 Hz, with identical 10-ms windows and 2.5-ms hops, five resonator channels outperform a single broadband RMS envelope by 21.1-44.9 percentage points across the tested architectures. The authors caution that different training budgets limit attributing this gain to spectral selectivity, and that the comparison does not establish an advantage over an equivalent five-channel digital filter bank.
What the measured envelopes preserve
Measured class templates averaged over 50 repetitions per class show that digits differ both in which channels respond and in when those responses occur. For "four," the response contains a sustained mid-utterance burst with strong R300, R1800, R4500, and R7500 responses and a weaker R3500 response. For "eight," an early burst is concentrated in R3500-R4500 in both speakers, with a separate late high-frequency burst more visible for speaker 22. Because this timing structure would be lost by pooling each channel over the utterance, the pipeline retains a temporal network to use it.
The paper also flags what these templates do not show. A resonator continues to respond after excitation ends, and the 10-ms RMS integration further smooths its output, so ringing and integration can blur the onsets and transitions that distinguish short speech sounds. The envelope integration interval is shorter than the baseline's 50-ms log-Mel window, with both using a 10-ms hop, but the authors state this difference alone does not establish the array's temporal resolution because the ring-down of each fabricated channel was not measured. The separated bursts remain distinguishable in the averaged envelopes; they do not provide a measured bound on transient response.
What the comparison does not isolate
The paired models retain each classifier family's main structure but differ in input representation and architecture, so the comparison measures the complete pipeline tradeoff rather than isolating the effect of physical filtering alone. The channel bank was selected on alphabet speech, and the narrow selection margin means the five-channel choice should be read as the highest-scoring bank tested, not as a proven optimum.
The scope is also bounded by the task and the hardware. Results come from spoken-digit recognition on a speaker-disjoint AudioMNIST split, with a prototype resonator bank and a Raspberry Pi 4 running one thread at batch size 1. The authors list integrating sensing and inference hardware, measuring total system energy, and evaluating continuous speech as future work. The reported energy figures therefore cover processor computation only, and the accuracy figures apply to this digit benchmark rather than to speech recognition generally.
The threshold that matters most
The most decision-relevant split in this paper is not the headline 98.6% reduction but the 90%-accuracy-in-under-10-ms threshold. Five Lyric configurations clear it and no evaluated log-Mel configuration does, which is where the representation changes what a small device can attempt. If the target is the highest accuracy in the comparison, the paired EdgeSpeechNet-A result is the relevant one: 95.70% at 6.56 ms and 1.97 mJ, a 3.58-point drop from 99.28% for roughly 69 times less compute time and estimated processor energy.
If the target is the cheapest model above 90% accuracy, the relevant comparison is PoolCNN at 91.07% and 1.81 ms against log-Mel TC-ResNet8 at 16.87 ms and 5.06 mJ. The paper's own framing supports reading these as a tradeoff curve rather than a single verdict: the gain is largest where latency and processor energy dominate, and it narrows when the baseline is already a small log-Mel model.
Sources used in this answer
Lyric: Wave-Domain Computing for Efficient Spoken-Digit Recognition
Lyric extracts speech features with passive acoustic resonators before digitization, acquiring 32 times fewer scalar samples than a 16-kHz waveform and cutting preprocessing-plus-inference latency and estimated processor energy per inference by 98.6% versus EdgeSpeechNet-A on a Raspberry Pi 4, at a 3.58-percentage-point accuracy cost to 95.70%.