Memory Management and Multithreading for Low Latency and Stability in On Device Voice Assistants
Bibliographic record
Abstract
Intelligent voice assistants must keep delay low and service steady while running on devices with limited compute and power. We used a two-path design that combines fine memory control (fixed pools, pinned buffers, reuse) with multi-thread scheduling (separate I/O, feature, and decode threads with bounded queues and deadlines). A total of 128 trials were run on a smartphone SoC, an embedded Linux board, and a wearable-class microcontroller in both booth and office settings. Power on the main rail was sampled and aligned with software logs, and idle energy was subtracted. Median latency fell by about 37% and p95 latency by about 38–40%. Jitter dropped by roughly 25%, and timeout events fell by about 74%. Throughput increased by 21–28%. Memory pools and pinned buffers shortened allocation time by about 61–63% and reduced DRAM traffic by about 31–34%, improving the energy–delay product by 18–24%. Recognition accuracy stayed stable (WER ≤ 0.2 absolute; F1 change ≤ 0.3). These results indicate that controlling memory use together with task priority can reduce delay and improve stability without loss of accuracy. The method is applicable to phones, wearables, and embedded boards, but the study is limited by a small set of devices, short test periods, and English-only data. Future work will add far-field and multilingual speech, longer trials, and secure-execution tests.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.003 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".