Voice check
Every answer the call recorder captured is cut out and measured: sound quality by a speech-quality model, how much the pitch rises and falls, the pace, the pauses, and repeated phrases. The figures give a provisional grade that I confirm or change by listening.
- When it runs
- After every recorded call
- Used in
- 16 reviews
- Runs on
- The call recordings, then my ears
What the voice check is
Natural is a judgement, but the parts of it can be measured, and measuring them first stops the grade from drifting with my mood. The check scores each answer on sound quality, intonation and delivery, combines them into a provisional grade, and the listening then confirms or overrides it. Accent is never part of the grade.
How it works
- It takes the companion's side of each call recording and cuts it into separate answers.
- Each answer is transcribed, so a garbled or cut-off answer is visible, and its length and words per second are counted.
- A speech-quality model rates the sound of each answer: the voice itself, the background and the overall impression, on a one-to-five scale used in speech research.
- A pitch tracker measures how much the voice rises and falls, which is what makes a voice sound spoken rather than read.
- Loudness variation, pauses inside the answer and repeated three-word phrases are counted.
- A fixed rule turns the pitch spread into an intonation score, docks the delivery score for an odd pace, repetition or a long pause, and combines sound quality, intonation and delivery into a provisional grade.
- I listen to the answers and keep or change the grade; the review says which.
What it records
- Per answer
- Its file and timing, the transcript, the three sound-quality figures, the pitch mean, spread and range, loudness variation, pauses, words per second, repeated phrases, and the provisional grade.
What it cannot catch
- The figures describe the sound, not whether the emotion fits the words; that part is listening only.
- Very short answers transcribe poorly and give weak pitch figures, so apps that answer in a few words are graded more by ear.
- Most voice grades are marked provisional until more calls are recorded.
Where it was used
Every published review that ran this checker, with the figure it produced there, read from the review’s own data. Each figure opens the test results behind it.
| Review | Figure | Tested |
|---|---|---|
| 4 (ringback excluded) Answers scored | 13 Oct 2026 | |
| 3.32 of 5 Speech quality | 18 Oct 2026 | |
| test results | 18 Oct 2026 | |
| test results | Invalid Date | |
| 2.9 to 3.1 of 5 Speech quality | 19 Oct 2026 | |
| 3.27 / 2.70 / 3.30 of 5 Speech quality (Rae / Moira / Yuna) | 15 Oct 2026 | |
| 2.7 to 2.9 of 5 Speech quality | 16 Oct 2026 | |
| 49 (Tania 17, Layan 16, Nadia 16) plus 1 voice note Answers analysed | 14 Oct 2026 | |
| 3.32 / 3.31 / 2.87 of 5 Speech quality (Luna / Priya / Gloria) | 15 Oct 2026 | |
| 44 (Ivy 16, Diana 16, Rin 12) Answers analysed | 14 Oct 2026 | |
| 3.31 / 3.41 / 3.45 of 5 Speech quality (Allie / Natalie / Marisol) | 15 Oct 2026 | |
| 3.41 of 5 Speech quality (Hana) | 15 Oct 2026 | |
| 2.5 to 2.8 of 5 Speech quality | 16 Oct 2026 | |
| 3 replies of 18 to 20 s Spoken replies (19 Oct) | 14 Oct 2026 | |
| 2.9 to 3.3 of 5 Speech quality | 18 Oct 2026 | |
| 4 voices, 1.0 to 1.3 s each Samples analysed | 14 Oct 2026 |
Review
- Tested
- 13 Oct 2026
- Figure
- 3.32 of 5 Speech quality
- Tested
- 18 Oct 2026
- Figure
- test results
- Tested
- 18 Oct 2026
- Figure
- test results
- Tested
- Invalid Date
- Tested
- 19 Oct 2026
- Tested
- 15 Oct 2026
- Tested
- 16 Oct 2026
- Tested
- 14 Oct 2026
- Tested
- 15 Oct 2026
- Tested
- 14 Oct 2026
- Tested
- 15 Oct 2026
- Tested
- 15 Oct 2026
- Tested
- 16 Oct 2026
- Tested
- 14 Oct 2026
- Tested
- 18 Oct 2026
- Tested
- 14 Oct 2026
The other checkers for ai girlfriend
Written by Miles Foster, Owner and tester, RankDude. How RankDude tests and grades is on the methodology page.