Blog

Research and notes from the team building Better Translate. Also see how it compares.

• Conversation runs on milliseconds. Passing a phone takes seconds.

  Every language studied minimises the silence between turns. Handing a phone across a table inserts several seconds into a system tuned that tightly, which is why translated conversations feel broken even when the translation is right.

• What interpreters know about where to sit

  Research on interpreted medical consultations keeps arriving at the same finding: where the interpreter sits decides who looks at whom. A phone on a table is making the same decision, usually badly.

• Why translated conversations feel awkward even when the translation is right

  People blame the translation. Usually the translation was fine and the interaction was broken: the timing, the gaze, and who is holding the device.

• The latency budget of a translated conversation

  Where the seconds actually go when two people talk through a phone, and which of them are ours to spend. Transcription is on the device and costs nothing on the wire.

• We ran 32 models over the same Japanese menu

  Our own menu benchmark, in full. Listing more languages does nothing, automatic detection is free, and the recogniser sets a ceiling no model can climb past.

• What context actually means on a menu

  Context is not a marketing word here. It is the section heading three lines above the dish, and it is the difference between autumn sword fish and Pacific saury.

• The slower model returned less of the menu

  One config spent 1,029 seconds on a photographed izakaya board and came back with less of it than another returned in 94. Waiting longer is not the same as thinking harder.

• The failure that looks like success

  One config kept 93% of the toppings and course components on our test menus. Another kept 22%. The second one produces a menu that reads perfectly and is quietly missing things you could have ordered.

• Thinking harder only helps a bad photograph

  Going from low to medium effort bought +0.08 exact match for twice the wall clock. Medium to high bought +0.01 for another 2.6 times. On a clean menu all three returned the same thing.

• The French menu that taught us not to assume

  One setting took a French set menu from 0.14 recall to 0.57 and its confidence from 0.38 to 0.98. The bug was that we told the phone what language to expect.

• We test on photographs, not on text files

  It is much easier to benchmark a translator by feeding it clean text. It also tells you almost nothing about whether somebody can order dinner.

• How to photograph a menu so it actually reads

  The same izakaya board recalled 0.63 of its dishes at night and 0.32 in daylight. The photograph moves the result further than anything on our end does. Four things that help.

• A thin scan is more dangerous than a wrong one

  A mistranslated dish announces itself. A dish that never appears does not. Here is how to spot a menu that came back short, and what we do to make it visible.

• The Japanese menu headings worth knowing before you sit down

  Five headings that organise most izakaya menus, taken from the boards we test against. Knowing the heading tells you more than translating a single dish does.

• Otoshi: the charge on your bill nobody warns you about

  お通し is a small dish you did not order and a seat charge you cannot refuse. It appears on our own test menu at 300 yen. Here is what it is and why it is not a scam.

• Do not use any translation app for allergies

  Ours included. A dropped negation turns a warning into its opposite, fluently, with nothing on screen to suggest anything went wrong.

• The three ways machine translation fails, and the one that matters at dinner

  Mistranslation, hallucination and deletion. Research names all three, and deletion is the one that quietly drops a negation, which is how an allergy warning becomes its opposite.

• Why the far half of the screen is upside down

  The obvious design is one screen you take turns with. We rotated half of it 180 degrees instead. Here is what that fixes and the one thing it costs.

• One microphone, one side at a time

  Both halves of the screen are live. Only one of them can listen. That is a hardware fact rather than a design choice, and pretending otherwise would just produce two bad transcripts.

• Why there is no account

  No sign-up, no email, no password, no profile. Nothing you translate is attached to a person because there is no person on file. Here is how that works with a paid app.

• Why the audio never leaves your phone

  Speech is transcribed on the device by Apple's SpeechAnalyzer, which has no server path at all. Only the resulting text is sent, and here is what that costs and what it buys.

• What it costs us to translate a menu

  Measured, not estimated: $0.0059 per menu page and $0.000147 per conversation turn. Here is the arithmetic behind a $23.99 year, and why the numbers are on a public page.

• Why there is no offline mode

  Transcription works on the phone with no connection. Translation does not, and will not soon. Here is the honest reason, and what to do on a plane.

• What ~600 reviews of a top voice translator actually complain about

  We read about 600 Play Store reviews of a leading voice translation app and counted the complaints. The recurring ones are mostly not about translation quality at all.