Why we built this benchmark
Every “best dictation app” list on the web ranks software its author has used casually for a week. Impressions vary by microphone, accent, and subject matter. The same app can feel 90% accurate in a quiet home office and 60% in a coffee shop. Without controlled conditions, rankings are anecdote.
We set out to fix that. Voice-list runs every app through the same five audio samples, on the same machine, in the same acoustic environment, every time. Numbers are numbers — you can disagree with our weights but not our measurements.
The audio samples
We use five samples that cover the main dictation contexts:
| ID | Context | Duration | Key challenge |
|---|---|---|---|
| en-01-coding | Developer dictation | 2 min | Technical jargon, identifiers, punctuation |
| en-02-casual | Casual conversation | 3 min | Contractions, filler words, informal grammar |
| en-03-conference | Formal meeting speech | 4 min | Proper nouns, numbers, fast speech |
| en-04-longform | Narrated prose | 6 min | Sentence variety, paragraphs, rare vocabulary |
| en-05-numbers | Numbers & data | 1 min | Dates, prices, percentages, phone numbers |
All samples are recorded by the same speaker in a treated acoustic environment at 44.1 kHz stereo, then down-sampled to 16 kHz mono (Whisper’s native format) before submission to each app. Ground-truth transcripts are verified by hand against the original recording.
The three error metrics
Word Error Rate (WER)
WER is the standard metric for automatic speech recognition:
WER = (S + D + I) / N × 100%Where S = substitutions, D = deletions, I = insertions, N = total words in reference. A substitution is a wrong word, a deletion is a missing word, an insertion is an added word that wasn’t spoken.
Lower is better. We mark WER bands as:
- Excellent — under 5% (near-human accuracy)
- Good — 5–10%
- Fair — 10–20%
- Poor — over 20%
Character Error Rate (CER)
CER applies the same formula at the character level rather than the word level. It is more sensitive to small typos and partial word errors. An app that consistently mishears the last syllable of words can have a WER of 10% but a CER of only 4%, because most characters are still correct.
We report both. Apps with very different WER and CER are typically strong on word structure but weak on specific phonemes or vice versa.
Punctuation Error Rate (PER)
PER measures errors on punctuation tokens only: periods, commas, question marks, colons, and capitalisation of sentence-initial words. We isolate these from the overall WER score because some apps produce “raw” transcripts (no punctuation added) while others add it via a second-pass LLM.
Why PER matters: A transcript like “the meeting starts at 9 am on thursday next week we need to confirm the budget” requires manual clean-up before pasting into a document. An app that outputs “The meeting starts at 9 AM on Thursday. Next week we need to confirm the budget.” saves real editing time.
Recording procedure
- Boot a clean session (no other apps running)
- Set the default input device to our reference microphone
- Start the app and configure it to its recommended default settings
- Read the sample aloud from the printed script (no deviations)
- Stop dictation, export or copy the raw transcript
- Record latency from end-of-speech to first character appearing on screen (warm) or first run after launch (cold)
- Repeat three times, take the median transcript for scoring
We do not cherry-pick the best run. The median run reflects typical production use, not ideal conditions.
The five scoring axes
Our overall 10-point score is a weighted average of five axes:
| Axis | Weight | What we measure |
|---|---|---|
| Accuracy | 35% | WER + CER + PER average across all 5 samples |
| Speed | 20% | Warm latency (ms from end of speech to text) |
| Privacy | 20% | Where audio goes: local = 10/10, cloud = scored by policy |
| Features | 15% | Hotkeys, modes, LLM post-processing, offline capability |
| UX | 10% | Onboarding, overlay design, editing workflow |
The weights reflect what users report as most important in our informal feedback collection. Speed and features are secondary to accuracy and privacy for professional dictation users.
What we do not test
- Non-English languages — our current samples are English only. We plan Russian, Spanish, and French samples in a later phase.
- iOS and Android apps — this site focuses on desktop dictation. Mobile behaviour may differ significantly from the desktop builds.
- Microphone sensitivity — we use a consistent reference microphone. Results will vary with budget headsets or built-in laptop mics.
Transparency
Every test run date is shown on each review page. If an app changes its
backend or ships a new model, we retest and update the tested_at date.
We do not accept payment for reviews or rankings. See our
editorial policy for full conflict-of-interest disclosures.