Original measurement log · Issue 02
Which voicemail details survive transcription?
We recorded one fictional callback card 12 ways, then checked whether a fixed speech-to-text model recovered the caller’s name, phone, ZIP code, and job intent exactly. Listen to every file and inspect every machine transcript below.
Direct answer
The recognizer recovered 41 of 48 scored fields exactly across this 12-file synthetic set. Phone and intent survived in all 12 files; the caller name survived in 7 and the ZIP code in 10. That result is evidence about these files—not proof that faster speech, repetition, or added noise generally helps.
The two-noise comparison was not monotonic: the harsher 5 dB files scored 7/8 fields while the 20 dB files scored 6/8. With only two voices per condition, that is a warning against turning a tiny test into a universal rule.
41/48
exact callback fields
7/12
files with all four fields
12/12
phone numbers exact
7/12
caller names exact
Fixed callback card
Same facts, controlled changes.
Only wording, voice, pace, or added white noise changed. The identity and callback details are deliberately fictional.
- Caller name
- Maya Chen
- Phone
- 415-555-0198
- ZIP code
- 94107
- Intent
- water heater estimate
Condition snapshot
Descriptive scores, not a leaderboard.
Each tile contains only two files, one per synthetic voice. Use it to navigate the recordings, not to rank writing techniques. Repetition preserved the phone again, but it did not repair the name error in the David sample.
Clean, neutral, two voices
Compact wording
6/8
exact fields across 2 files (75%)
Clean, neutral, two voices
Grouped wording
7/8
exact fields across 2 files (88%)
Clean, neutral, two voices
Repeated phone
7/8
exact fields across 2 files (88%)
Clean, fast, two voices
Faster grouped read
8/8
exact fields across 2 files (100%)
Grouped, neutral, two voices
20 dB noise
6/8
exact fields across 2 files (75%)
Grouped, neutral, two voices
5 dB noise
7/8
exact fields across 2 files (88%)
Research data use
The covered compiled result tables and measurements, results JSON, and paired-wording analysis CSV are provided under Call Safety Guide Research Data Terms v1. Download and reanalyse the covered results, or quote limited results with attribution to Call Safety Guide and a link to this lab. Redistribution of complete or modified datasets and commercial repackaging require prior written authorization.
Synthetic WAV/VTT audio and caption files and third-party assets are excluded; permission to reuse them has not been established. Automatic transcripts and file links are audit information, not a grant of third-party rights or permission to reuse referenced assets. Business use of greeting scripts remains governed by the existing script terms.
Field recovery
Numbers were not equally fragile.
| Field | Exact files | Observed miss | What an owner should do |
|---|---|---|---|
| Name | 7/12 | “Chen” became “Chan” or “Achan” in five files. | Ask callers to spell an unfamiliar name when it matters. |
| Phone | 12/12 | No exact misses in this controlled set. | Still repeat the number back during the callback. |
| ZIP | 10/12 | “94107” became “9407” in two files. | Confirm location before promising service coverage. |
| Intent | 12/12 | No exact phrase misses in this controlled set. | Use a short, concrete job phrase rather than a long story. |
Paired analysis · same synthetic voice
Did extra wording buy more correct details?
These four comparisons use only the six clean, neutral-speed files. Each row holds the voice constant and compares two scripts. Times are differences between the published durations rounded to one decimal place. This is a new analysis of the September 4 recordings, not a new recording session.
| Voice and wording change | Duration | Extra time | Exact fields |
|---|---|---|---|
| Zira: compact → grouped | 12.5s → 15.0s | +2.5s | 4/4 → 4/4 |
| Zira: grouped → repeated | 15.0s → 21.2s | +6.2s | 4/4 → 4/4 |
| David: compact → grouped | 12.6s → 15.1s | +2.5s | 2/4 → 3/4 |
| David: grouped → repeated | 15.1s → 21.2s | +6.1s | 3/4 → 3/4 |
Repeating the phone added time, with no score gain here.
The repeated scripts added 6.2 seconds for Zira and 6.1 for David over the grouped scripts. Their exact-field counts stayed at 4/4 and 3/4 respectively. Both baseline phone numbers were already correct, so the phone score had no room to improve. This comparison cannot show whether repetition helps when a phone number is initially missed, or when a human is listening.
One ZIP recovered; the name still needed checking.
David’s compact transcript contained “Maya Chan” and “9407”. The grouped version contained “Maya Chan” and “94107”: one more exact field for 2.5 extra seconds. Zira’s score stayed at 4/4. The scripts change wording as well as punctuation, so this does not isolate pauses or digit grouping as the cause.
Turn the observation into a verification step.
For these fictional messages, a callback handler should treat the name as unconfirmed and check the ZIP before deciding service coverage. A possible follow-up is: “Could you spell your last name and confirm your five-digit ZIP code?” That wording is our editorial recommendation; its effect was not tested in these recordings.
A useful next experiment would include multiple fictional names and ZIPs, a spelled-name variant, repeated runs and real phone-system recordings. Until those data exist, we cannot recommend a universal best script or claim a reduction in missed bookings.
All audio and machine transcripts
Audit every scored file.
A green mark means the target field appeared exactly after simple case, punctuation, and digit normalization. Word error rate is a diagnostic for this fixed transcript, not a usability score.
Sample 01 · Zira
compact · neutral · clean
4/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. Call me at four one five five five five zero one nine eight. My ZIP is nine four one zero seven. I need a water heater estimate.
Machine transcript
Hi, this is Maya Chen. Call me at 4155550198. My zip is 94107. I need a water heater estimate.
Sample 02 · David
compact · neutral · clean
2/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. Call me at four one five five five five zero one nine eight. My ZIP is nine four one zero seven. I need a water heater estimate.
Machine transcript
Hi, this is Maya Chan. Call me at 4155550198. My zip is 9407. I need a water heater estimate.
Sample 03 · Zira
grouped · neutral · clean
4/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chen. My callback number is 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 04 · David
grouped · neutral · clean
3/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chan. My callback number is 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 05 · Zira
repeated · neutral · clean
4/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. I will repeat that: four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chen. My callback number is 415-555-0198. I will repeat that, 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 06 · David
repeated · neutral · clean
3/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. I will repeat that: four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chan. My callback number is 415-555-0198. I will repeat that, 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 07 · Zira
grouped · fast · clean
4/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chen. My callback number is 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 08 · David
grouped · fast · clean
4/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chen. My callback number is 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 09 · Zira
grouped · neutral · 20 dB SNR
4/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chen. My callback number is 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 10 · Zira
grouped · neutral · 5 dB SNR
4/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chen. My callback number is 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Sample 11 · David
grouped · neutral · 20 dB SNR
2/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is my Achan. My callback number is 415-555-0198. My zip code is 9407. I am calling about a water heater estimate.
Sample 12 · David
grouped · neutral · 5 dB SNR
3/4 exact
Compare source and machine transcript
Source
Hi, this is Maya Chen. My callback number is four one five, five five five, zero one nine eight. My ZIP code is nine four one zero seven. I am calling about a water heater estimate.
Machine transcript
Hi, this is Maya Chan. My callback number is 415-555-0198. My zip code is 94107. I am calling about a water heater estimate.
Reproducible method
A narrow test with the settings exposed.
The generated files and scored JSON are published so the result can be checked without trusting a screenshot or summary claim.
Voices and format
Microsoft Zira Desktop (en-US) and Microsoft David Desktop (en-US). Files were normalized to 16 kHz, 16-bit, mono WAV.
Controlled conditions
Three wording versions; neutral and +2 SAPI rate; clean, 20 dB SNR, and 5 dB SNR white-noise files.
Transcriber
Systran/faster-whisper-small, faster-whisper 1.2.1, CPU int8, beam size 5, English fixed.
Deterministic noise
White-noise generation used the fixed seed 20260904; no sample was swapped after scoring.
Limits before conclusions
This is not a phone-network or human-caller study.
- This is a controlled synthetic-voice test, not a study of human callers.
- It does not reproduce carrier compression, voicemail platform processing, room echo, accents, stress, or background speech.
- Results apply only to these 12 files, this recognizer version, and the disclosed settings.
- Field scoring checks exact target recovery; it does not measure listener comprehension or business conversion.
Practical next step
Design the callback process as if transcription can be wrong.
Ask for a name, callback number, service area, and one concrete reason for calling. Then verify identity and location during the return call. A transcript can speed triage, but it should not be the only record used to promise coverage, price, or arrival.