Kokekun · measured comparison · 15 October 2026

V2, v3 and v3.5.
The same inputs, side by side.

100 matched trials attempted across 50 distinct inputs. Plan: 100 trials per version, 50 inputs repeated twice across 21 families. Both language directions; typed text and saved recordings. Original 20 plus resumed 80; no replay.

COMPLETE300 planned runsTwo measurement segmentsCode unchanged · daily allowance increased

First audio means the first received MP3 bytes, which may precede the first decodable frame. These measurements do not include microphone capture, Bluetooth delivery, audible onset, or a pronunciation listening score.

One planned study, completed in two segments. The original 20 matched trials were preserved, followed by the remaining 80. Between them, the authorized daily allowance changed from 2,000 to 3,000; deployed code module hashes and other Worker settings were verified unchanged. No completed jobs were replayed, and usage counters were not reset.

What these measurements show

V3.5 returned audio for 100/100 attempts, with guidance on all 50/50 Japanese outputs. V2 returned 17 Japanese clips without guidance. V3 had 16 guidance failures and 1 temporary service response (503) on English output.
Across all 100 matched trials, v3.5 received its first audio bytes a median 126 ms sooner than V2 (95% family interval: 53–275 ms sooner).
On the 83 trials with complete audio from all three versions, v3.5’s first received bytes arrived a median 68 ms later than v3 (interval: 0.46–177.5 ms later), while its final audio bytes arrived 1,198 ms sooner (interval: 512–1,796 ms sooner). These are paired changes, not differences between separate medians.
Within 2 seconds, the first bytes of clips that ultimately completed had arrived for 80/100 V2, 76/100 v3 and 94/100 v3.5 attempts. Failed attempts remain in these denominators.
The faster startup version switched between v3.5 and V2 across repeats for 26/50 inputs. This does not yet support a stable rule for choosing a version from the input.

The intervals describe this corpus and the two observed measurement segments. Received bytes are not audible onset; these results do not establish pronunciation accuracy or listening quality.

Availability first, then speed

Guided Japanese audio is reported separately: the V2 site can speak without guidance, while v3 and v3.5 require it. A returned clip is not proof that its pronunciation or meaning is correct.

Version	Audio returned	Guided Japanese	First audio · median	First audio · p90	Last bytes · median	Last bytes · p90
V2	100/100	33/50	1,131 ms	2,767 ms	5.22 s	10.66 s
V3	83/100	34/50	863 ms	1,679 ms	4.92 s	12.32 s
V3.5	100/100	50/50	960 ms	1,744 ms	3.45 s	9.68 s

Availability uses attempted denominators, with English → Japanese attempts counted separately. Unstarted planned trials are excluded. Reliability is measured across all 100 attempts per version in the completed study. Timing columns use only the same 83 successful trials on all three versions. P90 is the nearest-rank 90th percentile.

Audio availability, Japanese guidance, first-byte and complete-audio cumulative distributions, and matched successful timing distributions — The cumulative curves retain failed attempts in the denominator. Curves ending below 100% indicate missing complete audio, not faster responses. Both cumulative time axes are linear through 1 second, then logarithmic; the timing distributions use logarithmic vertical axes so long delays remain visible. Export vector figure.

Matched gains and losses

Negative differences favor the first named version. Each row below uses all successful pairs for that comparison; the figure uses the common three-version subset. Confidence intervals resample complete scenario families 5,000 times, preserving related repetitions.

Candidate − reference	Matched pairs	Median paired change	95% family interval	Candidate faster
V3 − V2	83	-218 ms	-260 ms to -145 ms	65/83
V3.5 − V2	100	-126 ms	-275 ms to -53 ms	62/100
V3.5 − V3	83	68 ms	0 ms to 177 ms	32/83

Paired first-audio timing differences and startup stage durations — The stage chart uses means so the three stages add to the mean first-audio time. Recognition is zero for typed input. Speech preparation, provider startup and network delivery are included after translation. Export vector figure.

Do patterns persist across inputs?

The two groups below reuse the previous study’s input-only clustering. Their membership was fixed before this run; new performance results were not used to redefine them.

Input group	V2	V3	V3.5
Shorter / fewer constraints Frozen group 1 · 25 planned inputs	50/50 audio 8/20 guided Japanese	37/50 audio 8/20 guided Japanese	50/50 audio 20/20 guided Japanese
Longer / more constraints Frozen group 2 · 25 planned inputs	50/50 audio 25/30 guided Japanese	46/50 audio 26/30 guided Japanese	50/50 audio 30/30 guided Japanese

Paired startup change for every input and both repetitions, with fixed input groups, output direction and separate failure markers — Each row is one of the 50 original inputs; the two points are its repetitions. Lines crossing zero show a change in which version was faster. V2 may return unguided Japanese audio, so timing alone does not establish equivalent pronunciation behavior. Only complete audio pairs receive a timing point. A failed version is identified separately, so failures cannot appear as zero-time successes. Groups and output directions are fixed from input information. Some repeat pairs span the interruption; their change cannot be attributed to a single cause. Export vector figure.

V3 vs V2: the faster version switched between repeats for 16/41 inputs with two successful comparisons.
V3.5 vs V2: the faster version switched between repeats for 26/50 inputs with two successful comparisons.
V3.5 vs V3: the faster version switched between repeats for 22/41 inputs with two successful comparisons.

Two repetitions reveal variation but cannot establish a stable rule for choosing a version. Translated wording can differ between versions even when the source input is identical.

Explore and listen

Open a trial to compare its source, translations, guidance and saved clips at normal playback speed. Listening has not been scored. V3 and v3.5 stop speech when they cannot prepare the required reading; V2 may continue without guidance.

Output languageInputInput groupMeasurement segmentFind an outcomeSearch

What changed between the versions?

Behavior	V2	V3	V3.5
Default request path	Separate recognition, translation and speech requests	Persistent server session	Same persistent server session as v3
Japanese preparation target	40 ms deadline; may continue without guidance	40 ms target; awaits guidance	40 ms target; awaits guidance
Required guidance	Optional selected rules	At least one valid rule	Guidance for every sentence and retained critical number/name readings
Pronunciation compiler	Original shared service	Original shared service	Separate repaired service
Temporary compiler failure	May continue without guidance	Speech stops	One retry after 100 ms before speech; unresolved failures stop
Unsupported reading	May generate ordinary pronunciation	Speech stops	Explicit reading or alignment message; speech stops

Shared configured providers: Groq whisper-large-v3, Cerebras gpt-oss-120b and Fish s2.1-pro. Voice label SophiaClone, 44.1 kHz MP3 and normal speech speed are retained. Configuration and guidance metadata do not certify naturalness.

Two measurement segments

The same planned case order continued after the interruption. The first segment contains 20 inputs from repeat 1. The second contains the remaining 30 inputs from repeat 1 and all 50 from repeat 2. The two segments therefore contain different input mixes; the summaries below are descriptive, not evidence that the allowance change improved or worsened speed.

Measurement segment	V2	V3	V3.5
Original cases 1–20 15 Oct 03:57:08 UTC to 15 Oct 04:03:27 UTC	20/20 audio 4/10 guided Japanese 1,025 ms median startup on 15 common successful trials	15/20 audio 5/10 guided Japanese 1,203 ms median startup on 15 common successful trials	20/20 audio 10/10 guided Japanese 913 ms median startup on 15 common successful trials
Resumed cases 21–100 15 Oct 04:14:25 UTC to 15 Oct 04:34:31 UTC	80/80 audio 29/40 guided Japanese 1,168 ms median startup on 68 common successful trials	68/80 audio 29/40 guided Japanese 845 ms median startup on 68 common successful trials	80/80 audio 40/40 guided Japanese 1,042 ms median startup on 68 common successful trials

Each segment’s timing median uses its own common three-version successful subset. The overall intervals retain each whole scenario family, including its observations in both segments. They describe these two observed periods; they do not estimate day-to-day service variation.

Method and evidence

Session routes may report preparation times of zero or near zero. Treat those as limited-resolution service telemetry, not evidence that guidance takes no time. The client’s first-audio and last-byte measurements remain the comparison measures.

The full plan balances six version-order permutations within input type and direction, then reverses each input’s order in a separately shuffled second round. All 100 planned trials and both repetitions are included in the completed report. One request chain runs at a time with no automatic job retries. V2 uses its normal HTTP sequence; v3 and v3.5 use established server sessions. Connection setup and pacing are recorded outside measured job timing.

Shared usage: 983 before the first segment → 1908 after the second. The daily allowance increased from 2,000 to 3,000 between cases 20 and 21, with approval. No counter reset and no replay of completed jobs. Other active users may contribute to the usage difference.

This is a comparison of deployed product paths during two measurement segments. It does not isolate compiler speed from transport, provider variation, caching or wording changes. No timings from earlier, separate benchmarks are pooled here; the original 20 trials from this same study are preserved with their source hashes. Continuous recording recovery and earbud performance are outside this test.

Raw responses · Analysis · Measurements CSV · Interpretation notes · Arithmetic verification · Audio integrity checks · Study plan · Versions before · Versions after

initial-20: 60 jobs · original results SHA-256 fb55536fa36ae05b33fa91f900a136839f900353be4a208e31f1cb2e9c006d0a
continuation-80: 240 jobs · original results SHA-256 3c227e55ad8ab93d994dd81ca79884d5ad83b4552d7c7a0e57c0870a8234d83f

Original measurement record SHA-256: 169ec5dbb8adb76fb1701e4414702ee6b62178b467c46c95de42627e31797c70