Speech recognition limits

For words the AI says wrong rather than hears wrong, see Pronunciation fixes. For menu items it can’t find, see Menu recognition issues.

The AI mishears specific words

Speech recognition picks the most likely word for what it hears. Words that are uncommon on restaurant calls, short, soft-sounding, or close to a more common word get swapped: “hostess” heard as “hostages”, “staff” as “sad”, “stews” as “shoes”, “chicken karahi” as “chicken teriyaki”.

• Add the words to your location’s Difficult Words list, which the speech engine gets on every call. See Difficult Words in Voice configuration. Maple support can also add them: send the word that was said and what the AI heard.
• Review call transcripts in the Maple portal regularly to spot patterns early.
• Keep your full menu synced and well categorized. When the AI knows you serve stews, it’s more likely to hear “stews”.
• Maple regularly ships transcription improvements, so a word you reported before may already be fixed.

Accents and speaking styles

The speech models are trained mostly on standard American English, so strong accents, very fast speech, mumbling or unusual phrasing lower accuracy and can lead to wrong orders.

• Tell Maple support about the pattern (for example, “many of our callers are native Spanish speakers”) so it can be prioritized, and send examples of calls that went wrong.
• Have staff with different accents call the AI to find gaps, and report specific failures.
• Maple has shipped improvements aimed at accents and varied speaking habits. If you had problems before, test again.
• Staff using the AI themselves should speak clearly and at a moderate pace.

Background noise

The AI processes all audio on the call, so a TV, other conversations, kitchen sounds or street noise can be taken as part of the caller’s request. Noise handling needs engineering work, not a setting change.

• When staff place test calls, do it somewhere quiet.
• If noise causes repeated problems, tell Maple support what kind of noise (for example “busy dining room” or “TV in background”).

Non-English menu item names

Recognition defaults to English. The AI understands English translations (“chicken and cream”) but may not recognize an authentic Spanish, Chinese, Italian or other pronunciation of an item such as “Pollo a la Crema”.

1. Add an English or phonetic description to the item in your POS. Maple syncs item descriptions from your POS and the AI reads them alongside the name. For example, “Burrito a la Mexicana” → description “Burrito Mexican”; “Pollo a la Crema” → “Chicken with cream sauce”. Where there’s no clean translation, write how a customer would say it in English.
2. Don’t rename the item in your POS, which puts duplicate or confusing names on kitchen screens. Use the description field.
3. Don’t make the change on Maple’s Menus page if your menu syncs from a POS: edits there are replaced on the next sync.
4. Turn on another language if many callers speak it. See Language settings in Voice configuration.
5. Call the AI and say items the way your customers do. Report any that still fail, with the item name and roughly how it was pronounced.

Transcripts full of errors

Widespread transcript errors usually come from the issues above combining: accents, noise, uncommon words and non-English names. Poor phone audio or very quiet callers also contribute.

• Send Maple support specific calls or transcript segments. Many reported issues are quick to fix once identified.
• Review recent transcripts: Maple has shipped general AI improvements and a faster model that improve accuracy.
• Combine fixes. With non-English menu items and varied accents, use the description workaround and report the accent patterns.

Report issues with specifics (exact words, kind of accent or noise), and re-test a week or two after reporting.