Where V3 gains time.
Where it loses time.
A matched comparison of the public V2 site and V3. Explore response time, audio completion, pronunciation guidance, and differences between repeated inputs.
Same input, two systems
Each point is one matched trial. Below the line means V3 was faster. Click a point to inspect its text, timing, and audio.
Both runs must have a measured value. Failures remain visible below; they are never plotted as zero. On smaller screens, scroll charts horizontally to keep labels readable; the tables provide the same exact timing values.
How often is it faster?
The share of all attempts that received the selected result by each time. A curve further left is faster; a lower final height means fewer results arrived.
The summary table compares matched observations only. The default curve keeps all attempts in its denominator. Observed p90 describes this sample, not a service guarantee.
Locate the gain or delay
Separate waiting for a transcript, translation, the first audio, and the remaining audio. These measured intervals help locate patterns; they do not by themselves identify the cause.
Each row uses pairs with both stage measurements, so sample sizes can differ. Stage medians should not be added. The transcription row includes saved recordings only. Typed inputs have no transcription stage. Pronunciation preparation is included in translation-to-first-audio, not an additional stage. Server-reported measurements use a different clock from received timing.
Which inputs behave alike?
Inputs are grouped by their recorded features, without using measured latency to choose groups. Compare performance afterward to look for patterns worth testing again.
Feature definitions, associations, and clustering limits
Does the advantage repeat?
The same distinct input was tested twice. Each line connects its first-audio difference in repeat 1 and repeat 2. Below zero is faster V3; crossing zero means the faster site changed.
Only inputs with both sites returning audio in both repeats appear as lines. Filters apply first. Two observations per input show variation, not a reliable estimate of its long-term distribution.
Every comparison is inspectable
Sort a column to find the largest gains, losses, or failures. Open a trial for exact text and saved audio.
| Direction / type | Audio result | Guidance · V2 / V3 |
|---|
Failed audio is reported as a failure, not a slow successful run. Translation may still be available. No pronunciation or semantic quality score is inferred from timing.
Speed and meaning are different checks
Recorded language observations complement the timing data. They are not a listening score or a certified translation-quality evaluation.
Read the saved review observations
What these results can tell us
Keep reliability, timing, and language quality separate when deciding what to improve.
Matched design
Timing definitions
- First audio bytes: request start to the first nonempty MP3 chunk. The chunk may begin with metadata; this is not a decoded frame or audible onset.
- Last audio bytes: request start to the last received audio bytes, not the time the voice finishes playing.
- Difference: V3 minus V2 for the same input and repeat. A negative value favors V3.
- Median and p90: calculated from available matched observations, with p90 using nearest rank.
Important limits
- These requests use fixed directions and do not test automatic language detection, microphone capture, actual earbud playback, Bluetooth delay, continuous conversation, or recovery after a connection drops.
- Shared service load, warming, caching, run order, and changed wording may affect timing. Repeat 2 also occurs later in the same session, so a repeat difference cannot be attributed to caching alone. The design does not isolate a single cause.
- Saved recordings are controlled inputs. They do not represent the full range of real speakers or noisy environments.
- Clustering and associations are exploratory. Input features can overlap; a cluster is not a validated class of performance.
- Repeated trials are not independent inputs. Confidence intervals in the saved analysis resample whole input families, retaining related inputs and their repeats together.
- Audio received is not a pronunciation or translation-quality rating. Exact text and clips remain available for review.