Éist
Get App

Éist research · snapshot 2026-09-15-v1

On-device TTS performance in Éist

Observed real-time factor (RTF) from production installations with analytics enabled. Analytics is enabled by default and users can turn it off in Settings. Lower is faster: 0.5 means synthesis took about half the generated audio duration. These are field observations from Éist, not controlled engine benchmarks.

Observation window

25 Aug–15 Sep 2026

Qualified device cohorts

25

Engine/platform health samples

100,574

Engine and platform results

Each installation contributes one typical value. Do not interpret configured concurrency as throughput.

Equal-installation engine and platform summaries
EnginePlatformMedian RTFp10–p90InstallationsHealth samples
Kokoroandroid3.482.326–5.258631,998
Kokoroios1.3250.868–2.322595,614
Kokoro Turboios0.4380.227–0.6721052,602
Piperandroid0.2470.115–0.78229711,193
Piperios0.1150.088–0.2329316,620
Supertonicandroid0.7910.385–2.50524718,865
Supertonicios0.3040.212–0.53124443,682

Device cohorts

A device row is published only with at least 100 active health samples and 10 installations. Android models are too fragmented to clear the installation floor in this snapshot, so Android appears only in the platform-level table above.

Published device cohorts filtered by engine, platform and device
Engine Device Observed OS Median RTF p10–p90 Installations Health samples
Piper iPhone18,2 iOS 26.6–27.0 0.09 0.088–0.103 11 193
Piper iPhone18,1 iOS 26.3–26.6.1 0.098 0.089–0.192 10 360
Piper iPhone17,2 iOS 26.5.2–26.6.2 0.101 0.095–0.21 11 234
Piper iPhone18,3 iOS 26.6–27.0 0.101 0.086–0.162 14 969
Piper iPhone15,2 iOS 18.5–26.6.2 0.102 0.093–0.317 15 222
Piper iPhone17,1 iOS 18.6.2–27.0 0.103 0.085–0.175 17 512
Piper iPhone16,1 iOS 18.7.8–26.6.1 0.106 0.088–0.174 15 373
Piper iPhone16,2 iOS 26.5–27.0 0.11 0.093–0.129 11 168
Piper iPhone17,3 iOS 26.3–27.0 0.115 0.095–0.186 27 2,399
Piper iPhone15,4 iOS 26.3.1–27.0 0.122 0.097–0.214 14 1,428
Piper iPhone15,3 iOS 26.6–27.0 0.123 0.092–0.238 12 138
Piper iPhone14,7 iOS 18.7.2–27.0 0.133 0.1–0.249 13 2,468
Piper iPhone14,5 iOS 18.5–26.6.2 0.153 0.113–0.274 21 399
Piper iPhone12,1 iOS 18.2–26.6.2 0.196 0.166–0.797 16 260
Supertonic iPhone18,1 iOS 26.3–27.0 0.247 0.21–0.459 12 2,624
Supertonic iPhone17,1 iOS 18.5–27.0 0.271 0.218–0.36 17 2,235
Supertonic iPhone18,3 iOS 26.5.2–27.0 0.278 0.227–0.38 24 1,074
Supertonic iPhone16,1 iOS 26.5–26.6.2 0.281 0.216–0.427 14 2,181
Supertonic iPhone17,3 iOS 26.3–26.6.2 0.283 0.22–0.466 14 1,958
Supertonic iPhone15,4 iOS 26.5–26.6.2 0.309 0.221–0.428 12 5,868
Supertonic iPhone16,2 iOS 26.5–26.6.2 0.324 0.21–0.5 12 2,942
Supertonic iPhone14,7 iOS 18.7.2–27.0 0.394 0.339–0.477 10 391
Supertonic iPhone14,5 iOS 18.5–26.6.2 0.415 0.334–0.679 14 1,756
Kokoro Turbo iPhone17,1 iOS 18.7.8–27.0 0.469 0.333–0.673 12 275
Kokoro Turbo iPhone16,1 iOS 26.5–26.6.2 0.5 0.379–0.75 10 285

Download this snapshot

Immutable, privacy-screened aggregate data. The JSON includes platform summaries; the CSV contains the visible device table.

Download CSVDownload JSONChecksums

Citation

Éist. “On-device TTS performance in Éist.” Snapshot 2026-09-15-v1, methodology 1.0.0. Updated 21 September 2026.

Methodology and limits

The app emits a periodic synthesis_health observation while real synthesis work is active. avg_rtf is cumulative generated-audio time divided by synthesis time inside the active coordinator. We first calculate an exact median for each installation and then calculate exact cohort percentiles across those installation summaries. A highly active listener therefore gets the same weight as a lighter listener.

The snapshot includes release builds, foreground playback, physical devices, approved app versions, and values where 0 < RTF ≤ 10. It excludes emulators, sideloaded and test builds, background observations, system TTS, opted-out installations, and sparse cohorts. No raw identifier, timestamp, book, voice choice, location, or person-level row is published.

This dataset describes Éist’s implementation and usage mix. It does not measure speech quality, battery, memory, thermals, or global device ownership. Active health samples are not individual synthesis calls. Configured concurrency is not observed utilisation, so this release deliberately makes no throughput claim.

Configuration IDs

  • Supertonic: upstream v3 model, Éist cache version 7.
  • Piper: VITS through sherpa-onnx 1.12.40, Éist cache version 5.
  • Kokoro: fp32 model through sherpa-onnx 1.12.40, Éist cache version 5.
  • Kokoro Turbo: iOS MLX/Metal bf16 pack v3, Éist cache version 3.

Changelog

2026-09-21: first reviewed publication. Android device-level rows withheld because no exact model reached 10 installations.