What "didn't recognize" usually means

When voice logging fails, it usually fails in one of three specific ways. The fix depends on which one happened, so it's worth knowing what to look for.

The estimate looks wildly wrong. You said "Greek yogurt with honey" and the app returned 800 calories. Almost always, the speech-to-text got the wrong words and the model gamely estimated something else.

The estimate is generic and probably under-counts. You said "a smoothie" and got 180 calories. Real smoothies range from 200 to 700 calories depending on what's in them. "A smoothie" without specifics gets a midpoint guess that's usually wrong for any specific smoothie.

The app says it can't estimate. Rare but it happens. The description was either incomprehensible to the speech-to-text or didn't contain enough food information to attempt a guess.

Each one has a different fix.

Fix 1: check the transcript before confirming

After you record a voice log, TrakMac shows you the transcript before sending it to the estimator. Read it before tapping confirm. If the words don't match what you said, edit them.

The most common transcription errors:

  • Brand names that sound like other words. "Chipotle" sometimes becomes "chip butty." "Sweetgreen" sometimes becomes "sweet green" (two words). "Halo Top" can become "halo to." Most chains transcribe correctly most of the time, but the failure modes are predictable.
  • Numbers. "Two" and "too" and "to" all sound the same. "Two eggs" can become "to eggs." The estimate doesn't know what "to eggs" is.
  • Compound food names. "Beef bowl" can become "beef ball." "Stir fry" can become "star fry."
  • Background noise. Logging while a TV is on, while in the car with windows down, or in a loud restaurant cuts transcription accuracy substantially.

The edit field on the transcript screen is just a text input. Tap it, fix the words, hit confirm. The estimator runs on the corrected text.

Fix 2: be more specific

Vague descriptions get vague estimates. The estimator does its best with what it has, but "a sandwich" can be 250 calories or 900. The model defaults to the average for "sandwich," which is wrong for any specific one.

Descriptions that produce noticeably better estimates:

  • Include the protein. Not "a salad" — "a chicken Caesar salad." The protein is the calorie anchor for most meals.
  • Include the brand for restaurant food. Not "a burrito" — "a Chipotle steak burrito with rice and beans." TrakMac has a cache of menu items for the major chains; the brand triggers a more accurate lookup.
  • Include preparation method. "Grilled chicken" vs "fried chicken" vs "breaded chicken" are three meaningfully different estimates. Cooking method matters as much as portion size in many cases.
  • Include rough portion size when it's not standard. "A big bowl of pasta" or "about two cups of rice" gives the model something to work with. Without portion info it assumes a typical serving, which can be off by 50% in either direction.
  • Include condiments and dressings if they're significant. "With ranch" or "with mayo" can add 150-300 calories. The model assumes plain unless told otherwise.

A practical example: "a salad" becomes "a Sweetgreen Harvest Bowl with chicken, no goat cheese, balsamic dressing." The first version returns a guess. The second returns something close to the chain's published nutrition.

Fix 3: handle the genuinely uncommon

If the food is regional, niche, homemade with a non-standard recipe, or otherwise outside what the model would have seen often, the estimate may be a stretch even with good description.

For these cases:

Estimate the components. "A bowl of my grandmother's stew" is hard. "Beef stew with potatoes, carrots, onions, about 6 oz of beef and 1.5 cups of vegetables" is much easier. Break the dish into known parts.

Use a comparable. "Like a thick chili with sweet potato instead of beans, about 2 cups" gives the estimator a reference point. Most foods have a close-enough comparable in mainstream nutrition data.

Pull the recipe nutrition manually. If the dish is something you make from a specific recipe, the recipe blog usually has the nutrition info. Log the meal as a custom entry with those numbers (see editing a logged meal) instead of trying to talk through it.

When the estimator returns nothing

A few cases where you'll get an error or a blank estimate instead of a number:

  • The transcript was completely garbled (loud environment, very fast speech, microphone covered). Re-record in a quieter setting.
  • The transcript wasn't food (you accidentally hit the mic and spoke about something else). Cancel the entry; the app won't make up a meal.
  • Network failure during the estimate request. The transcript is local but the estimate runs server-side. If you're on a flaky connection, the request can time out. Try again on better network.
  • The model timed out on a very long, complex description. If you described 8 different items in one recording, the estimator can struggle. Break the meal into 2-3 separate logs.

What happens after a few corrections

The accuracy stack learns from your edits. The first time you log "my breakfast smoothie," the estimate is generic. The third time, after you've edited the macros twice, the estimator is calibrated to your typical smoothie composition and the estimate lands within ±10%.

This works per-user (your edits improve estimates for you specifically) and globally (when many users correct the same item the same way, the underlying model updates for everyone). The implication: editing isn't just fixing today's number, it's improving tomorrow's first guess.

If you have a meal you eat regularly that the estimator keeps getting wrong, the fastest path to a permanently accurate estimate is to log it 3-4 times and edit each time. By the fourth log, the estimate should land in the right zone without intervention.

When to stop fighting it and just type

If you've tried 2-3 voice descriptions of the same meal and the estimate is still wrong, switch to text. The text input on the same screen accepts the same description but skips the speech-to-text step, eliminating the most common source of error.

Voice is faster when it works. Text is more reliable when voice doesn't. Use the right tool for the meal in front of you.