What has actually been measured

On 22 August 2026, we submitted one 24.4-second synthetic recording to a local Journo backend connected to its real transcription and journal-processing providers. The same fixed file, script and evaluation format are prepared for six products.

ProductCurrent statusPublic evidence
JournoMeasured onceAudio, result JSON and response JSON published
AudioNotesNot measuredNone
VoicenotesNot measuredNone
CleftNot measuredNone
ChatGPTNot measuredNone
UntoldNot measuredNone

An empty cell would invite an assumption. “Not measured” means exactly that: we have not run a controlled test that meets the publication method below.

The current Journo baseline

MeasureResult
Recording length24.4 seconds
Time to readable server result9.655 seconds
Reference words82
Raw-transcript substitutions0
Raw-transcript deletions0
Raw-transcript insertions0
Journal-output word changes0
Journal-output formatting changesPunctuation and paragraph breaks

The run measured the server path. It did not measure phone upload time, offline capture, on-device audio retention, recovery after a force-quit, noisy speech or accent coverage. Those observations remain not_tested in the result rather than being inferred from Journo’s source code.

Download the evidence

The audio SHA-256 is a077ab55a51ca795cdf9e67087e3ceaa75a84b513ccfa7c6d7c2535957e6b443. The response-evidence SHA-256 is 2df9df2c6433d719203fab18e4d620fb081b243814c9cd6ac1b383f520193011. Those hashes let another reviewer verify that the files used for a rerun have not silently changed.

How the measurement works

  1. Use the identical audio file and record its SHA-256.
  2. Record product version, plan, device, network and UTC test time.
  3. Start timing immediately before submission and stop at the first complete readable result.
  4. Preserve the raw transcript and journal output exactly as returned.
  5. Compare lowercase word tokens with the fixed script to count substitutions, deletions and insertions.
  6. Publish the response evidence and its cryptographic hash.
  7. Mark anything not directly observed as not tested.

This first baseline used a synthetic voice and an ordinary, non-sensitive reflection so the artifact can be public. Real journal content must never enter the corpus.

What would justify a broader claim

This page will not name a speed, accuracy or reliability winner until all six products have been tested under the same method. A stronger publication requires, at minimum:

  • the fixed synthetic sample through every product;
  • three consenting human speakers with different speaking patterns;
  • identical timing and evidence capture;
  • product versions and test dates;
  • a rerun when a material product change makes a result stale.

Even then, the result would describe this corpus and protocol—not every speaker, language, room or device.

Reproduce or challenge it

Download the corpus and result. If you find a counting error, a stale product fact or a method that creates an unfair comparison, send the exact issue through Journo Support. Corrections and material method changes follow the public research methodology.

The useful outcome is not a flattering score for Journo. It is a test another person can inspect and a clear boundary around what the evidence does not show.

Sources and verification

We use first-party product documentation and original research wherever possible. Features and prices can change.

  1. Dated machine-readable voice-journal benchmark
  2. Fixed synthetic audio corpus
  3. Hash-verified Journo response evidence