Energy Use of Speech Algorithms on Intelligent Terminals

2.1. Study Area and Sample Description

We collected 100 test runs from smartphones, embedded boards, and wearable devices. Measurements were carried out in two environments. The first was a controlled acoustic room with a temperature of 24 ± 1 °C and relative humidity of 45–55%. The second was a normal office with background noise of 50–60 dB. Devices covered different hardware classes, including processors with NPUs and microcontrollers with limited memory. All devices were fully charged before testing to avoid the effect of battery discharge.

2.2. Experimental Design and Control Setup

Two groups were used. The control group ran baseline speech models in floating-point precision. The experimental group used compressed and pruned models with modified front-end settings. Each task was repeated 40 times per device to reduce random effects. A paired design allowed direct comparison between optimized and non-optimized pipelines. This design was chosen because earlier studies showed that memory access and model precision strongly influence energy in speech tasks.

2.3. Measurement Methods and Quality Control

Energy was recorded with a digital power analyzer at a sampling rate of 5 kHz, connected to the main supply line. Both idle and active states were measured. Net energy was calculated by subtracting idle power from task power. Average power was taken during the stable part of each task. Device temperature was monitored with an infrared sensor, and tests with surface temperature rise above 2 °C were removed. Instruments were calibrated before experiments, and duplicate runs were included at the start and end of each day to check consistency.

2.4. Data Processing and Model Equations

Data were processed in MATLAB. Total energy EEE was calculated as [20]:

$$E=\sum _{i=1}^n{P}_i\text{⋅}\mathsf{\Delta }t$$

where $P_i$ is the measured power at interval $\mathsf{\Delta }t$, and $n$ is the number of samples.Normalized energy per inference was then computed:

$$E_n={\displaystyle \frac{E}{N}}$$

where $N$ is the number of processed frames. A regression model was also used:

$$E_n=\alpha M+\beta F+\gamma$$

where $M$ is the number of parameters, FFF is the feature dimension, and $\alpha ,\beta ,\gamma$ are coefficients estimated by least squares.

2.5. Statistical Analysis

Results are reported as mean ± standard deviation. Paired ttt-tests were used when data met normality. If not, the Wilcoxon signed-rank test was applied. A value of $\mathrm{p}\text{<}0.05$ was considered significant. Residuals from regression models were checked for independence and equal variance. These steps ensured that both experimental and control data were statistically sound and repeatable.

3.1. Overall Energy and Accuracy

Across 100 runs, the optimized pipelines used less energy than the baseline in both tasks. For streaming ASR, median per-utterance energy fell from 0.94 J to 0.70 J (−25.4%, 95% CI: −22.7% to −27.8%). For always-on KWS, average standby power dropped from 6.0 mW to 4.1 mW (−31.7%, 95% CI: −28.5% to −34.6%). Word error rate on test-clean changed by ≤0.15 absolute, and KWS F1 decreased by ≤0.2 points. The stage-wise energy attribution for ASR is shown in Figure 1 and guided where we applied low-precision compute and kernel fusion. Similar stage sensitivity for streaming ASR has been reported in recent work that links beam-search confidence to energy savings on edge hardware [21].

3.2. Ablations and Main Drivers

Single-factor tests showed consistent effects. Layer-wise sub-8-bit quantization on memory-bound blocks reduced energy by 15.2% at fixed accuracy. Saliency-guided pruning added 8.7%. Front-end frame-rate subsampling (10 ms → 20 ms) saved 6.2% but increased KWS false rejects; a lower wake threshold restored F1 within 0.1 points of baseline. DVFS with operator fusion reduced a further 7.3% on NPUs. A linear model with two covariates—parameter count and feature size—explained most of the variance in normalized energy (${\mathrm{R}}^2=0.80$). These drivers are in line with reports that target acoustic-model compute and runtime control to cut power in streaming ASR [22].

3.3. Robustness Across Devices and Modes

Trends held on different types of devices. On a handset SoC with an integrated NPU, fused kernels with DVFS reduced the energy–delay product by 26.5% while keeping the surface temperature rise below 2 °C. On a wearable board, duty-cycled feature extraction reduced standby power by 28.9%. On an MCU-plus-accelerator, a two-stage KWS pipeline—binarized wake detector followed by a higher-bit command recognizer—kept false alarms below 0.5/h and lowered standby energy. This layout is shown in Figure 2 and is consistent with recent MDPI reports on resource-efficient, low-bit KWS engines for edge devices.

3.4. Comparison with Prior Work and Limits

Our end-to-end savings are close to those reported for confidence-guided ASR, which achieved ≈25–26% lower energy and time for acoustic-model evaluation with minor accuracy change; here we add rail-level measurements and per-stage attribution across devices. For KWS, our two-stage, low-bit design aligns with MDPI results showing binarized CNN pipelines and dedicated engines that cut resources while keeping real-time response. Limits remain: the device set was small, field sessions were short, and non-English and far-field speech were limited. Future work should extend long-duration listening, add multilingual and noisy scenes, and test the joint cost of secure boot and integrity checks with energy control. Related co-design studies on simplifying MFCC front ends also support moving more saving to the feature stage [23].