On-device transcription benchmarks: Parakeet TDT 0.6B v3 (Q8_0 GGUF) via NeMo-Speech.cpp
Collected 2026-10-25 on a Google Pixel 8a (Tensor G3: 1x Cortex-X3, 4x A715, 4x A510), Android 17 / API 37.
Phone plugged in, screen on, test process in the foreground cpuset. Thermal status stayed at 1 (light).
Debug APK; native code built with -O3 Release, CPU backend only, GGML_OPENMP off.

CPU builds (GGML_CPU_ARM_ARCH):
  portable = armv8-a (repository default)
  dotprod  = armv8.2-a+dotprod+fp16
  i8mm     = armv8.2-a+dotprod+fp16+i8mm
Verified by disassembling libggml-cpu.so: portable has 0 sdot instructions; dotprod has 1,139;
i8mm has 1,139 sdot and 260 smmla.

Rounds:
  a = portable, dotprod, i8mm in that order; thread sweep 2/4/6/8, then 3 live replays at 4 threads
  b = i8mm, dotprod, portable (reversed order); 4 threads only, 1 live replay each
  c = i8mm only; 3-thread benchmark, then live replays alternating 3 and 4 threads
45-second idle gap before every test invocation.

benchmark-runs.csv
  One row per complete-file inference of the bundled 11.0-second JFK WAV (16 kHz mono).
  init_ms: model load plus two preview-shape warmup calls, measured once per test invocation.
  inference_ms: wall time of one native nemo_speech_asr_recognize_f32 call (from JNI).
  rtf: inference_ms / audio duration. Below 1 = faster than the audio's length.
  Five calls per invocation on one loaded recognizer; the first call is included.

live-replay-runs.csv
  The same WAV fed through the app's live pipeline in 512-sample (32 ms) frames at real-time
  cadence, replacing the microphone. Same segmentation, mailbox, and inference worker as the app.
  final_ms: time from the last voiced frame to the final text of the last segment.
  queue_age_ms: diagnostic; the largest wait observed by the last processed request or capture frame.
  first_preview_ms: time from the start of the LAST segment that produced a preview to that preview.
    It does not describe the first segment, and is 0 when no preview was shown for the last segment.
    Treat it as a diagnostic, not a time-to-first-text measurement.
  wer: word error rate against the reference after removing case and punctuation.

Limits: one phone, one 11-second English clip, no microphone capture, no peak-memory, battery,
or sustained-thermal measurement. Not percentile latency.