Back to Blog
NLP9 min read1,880 words

Resolving Time Expressions: The Hidden Difficulty in Voice Scheduling

Why turning "agle mangalwar saade paanch baje" into an exact calendar slot is one of the hardest small problems in voice scheduling.

Abstract cover of overlapping amber clock arcs and calendar grid lines dissolving into a single highlighted cell

Almost every useful phone conversation eventually lands on a time. A clinic call ends with an appointment, a service call ends with a technician visit, a sales call ends with a callback. The caller says something ordinary, the agent says "done", and a row appears in a calendar. The distance between those two things is much larger than it looks.

People do not speak in timestamps. They speak in references: relative to today, to a festival, to when they get off work. Converting those into a single unambiguous slot is a well-studied task in natural language processing called temporal expression resolution, and in Indian voice conversations it carries difficulties that a general-purpose date parser was never built for.

Three Kinds of Time Expression

Spoken time references fall into three families, and each fails differently.

  • Absolute expressions name a point on the calendar directly: "on the twelfth of November", "on the second at eleven". These are easiest to resolve, but still incomplete, since a date without a year needs a year and "the second" needs a month.
  • Relative expressions are defined with respect to some other moment: "tomorrow", "next week", "in three days", "same time next month". These dominate real speech, and they are useless without an anchor.
  • Recurring expressions describe a pattern rather than a point: "every Monday", "alternate Tuesdays", "har shanivaar subah". These cannot be represented as a single timestamp at all, and a system that quietly collapses one into a single booking has lost information the caller expects to have been captured.

The Anchor Problem

Every relative expression is resolved against a reference time. Choosing the wrong one produces errors that are individually small and collectively serious, because an off-by-one-day booking is worse than no booking at all. Several anchors can disagree at once:

  • The caller's clock. What the caller means by "tomorrow" depends on their local date at the moment they speak.
  • The server's clock. Machines commonly work in UTC. Near midnight the UTC date and the Indian date differ, so "tomorrow" resolved against a UTC date can be a day early.
  • The conversation's start time. A call beginning at 11:52 pm and ending at 12:06 am contains two different values of "today". Anchoring every expression to the start of the call is usually more consistent than anchoring each one to the instant it was spoken.

The practical rule is to fix one anchor explicitly, carry it as data through the whole conversation, and never let an implicit system clock decide.

Time Zones, Calendars and Callers Abroad

India uses a single time zone, UTC+05:30, and does not observe daylight saving time. That removes an entire class of bug: there are no local hours that fail to exist, and none that occur twice. A domestic booking system can treat local time as a clean, continuous line. The complexity reappears at the edges:

  • Non-resident callers. Someone in Dubai, London or Toronto may say "tomorrow morning" meaning their morning, or may already be converting to Indian time in their head. Several of those regions observe daylight saving, so their offset from the Indian day shifts twice a year.
  • Working weeks and holidays. Saturday is a working day for many Indian businesses, some close on a different weekday, and public holidays vary by state. "The next working day" has no single national answer.

For a caller whose location is uncertain, it is safer to state the zone than to infer it. "Thursday at four in the evening, India time" costs three words and removes the ambiguity.

Indian Day Vocabulary

The symmetric day words

In Hindi, kal means both yesterday and tomorrow. Parso means both the day before yesterday and the day after tomorrow. In many dialects narso extends the pattern one day further, three days away in either direction, though its use is regional and much less frequent, so a system should be cautious about resolving it confidently.

Direction is carried by tense and context rather than by the word: "kal aaya tha" is past, "kal aaunga" is future. Booking intent is almost always forward-looking, which makes resolution easier in practice, but the safe design is to take the future reading and then say the actual date aloud.

Gujarati is friendlier here, because it usually marks direction on the word itself: aavti kaale for tomorrow and gai kaale for yesterday. A system that treats Gujarati "kaale" as equivalent to Hindi "kal" throws away a disambiguation the language already provided.

"Agle mangalwar" and parts of the day

"Next Tuesday" is ambiguous in English, and "agle mangalwar" inherits the same ambiguity. Spoken on a Monday it may mean tomorrow or the Tuesday eight days away, and the same speaker may not be consistent. There is no correct rule, only a reasonable default plus an explicit confirmation, and the words that resolve it are the date itself.

Subah (morning), dopahar (afternoon), shaam (evening) and raat (night) are not decorative. They are the main way callers disambiguate a bare clock time, and their boundaries are soft. "Shaam ko paanch baje" is unambiguously 5 pm. A resolver should treat these as constraints on the 12-hour choice rather than as independent values.

Clock Forms That Trip Systems Up

  • saade marks half past, as in saade paanch for 5:30 and saade saat for 7:30. It is used from three o'clock upward.
  • dhaai is 2:30 and dedh is 1:30. These are separate words, not "saade do" and "saade ek", which are not idiomatic. A pattern that only looks for "saade" plus a numeral will miss two of the most common half-hour appointment times in the working day.
  • sawa marks a quarter past: sawa paanch is 5:15.
  • paune marks a quarter to, and critically it shifts the hour. Paune paanch is 4:45, not 5:45. A resolver that reads the numeral and adds forty-five minutes produces an answer exactly one hour late.
  • baje marks the hour itself: paanch baje is five o'clock.

Gujarati uses closely related forms, and Hinglish mixes freely: "saade five" and "paune chhe" can appear in one conversation. The mirror-image problem, speaking these forms aloud, is covered in our post on text normalisation for Indian speech synthesis.

Then there is the ambiguity of "paanch baje" with no part-of-day word attached. Two candidates exist, and only context decides. Useful constraints, in rough order of reliability:

  1. Any part-of-day word in the same or an adjacent utterance.
  2. The business's operating hours. If a clinic runs from 9 am to 8 pm, the morning reading of "seven baje" may not even be bookable.
  3. Whether one candidate has already passed. Nobody books an appointment for two hours ago.
  4. The rest of the request. "Office se nikalke aaunga" points firmly at the evening.

When both readings remain plausible and bookable, asking is better than guessing. One short question costs far less than a missed appointment.

Vague, Cultural and Moving References

  • Soft ranges: "next week sometime", "month end", "koi bhi din". These resolve to an interval, not a point, and the honest handling is to propose specific options inside it rather than to invent one.
  • Event-relative references: "after Diwali", "before Eid", "after the exams". Some are calendar events, some are personal, some are neither.
  • Moving festivals. Diwali follows a lunisolar reckoning and falls on a different Gregorian date each year. Eid follows the Islamic lunar calendar, moves earlier each Gregorian year, and its exact date can depend on local sighting announcements. A festival table therefore needs maintenance, a year dimension, and a defined behaviour for dates it does not know.

A vague expression should stay vague until the caller makes it specific.

Normalising, Representing and Validating

Whatever extracts the time, the output should be reduced to a single normalised representation before anything else touches it, and the standard vocabulary is well established. ISO 8601 covers points in time, durations and intervals. The TIMEX3 annotation style from the TimeML tradition, used in temporal tagging research and in the well-known open-source taggers, adds what matters for speech: a type for date, time, duration or set; a value that may be deliberately underspecified; and an explicit anchor. Underspecification is a feature, because it lets a system record that the caller said "next week" without pretending they said Tuesday. Recurrence needs its own representation, and the recurrence rule format used in calendaring standards expresses "every alternate Tuesday until March" directly.

Rules, models and why both belong

Rule-based resolvers are precise, inspectable and deterministic. Given a clean phrase and an anchor they return the same answer every time, and when they are wrong the rule that misfired can be found. Their weakness is coverage: speech contains disfluencies, self-corrections and code-mixing that no grammar anticipated.

Language models handle that messiness far better, coping with "Tuesday, sorry, Wednesday, evening mein, saade saat ke aas paas" without a rule for it. Their weakness is the mirror image: they can produce a date that is well-formed and wrong, and perform arithmetic incorrectly while sounding certain.

The sensible arrangement is layered. Extraction can be tolerant, but the extracted value should be emitted in a strict normalised form, then recomputed and checked deterministically against the anchor and the business calendar. Arithmetic belongs in code; interpretation belongs to whatever handles language best.

Validating the resolved value

  • Is the datetime in the future, and by a plausible margin? A booking eleven months out from a walk-in clinic is more likely a year error than a real request.
  • Does it fall inside operating hours, on a working day, in a slot that is free?
  • Was the anchor stored alongside the value, so the resolution can be audited?
  • Is confidence low enough that clarification is warranted? Anything failing validation should trigger a question, not a silent correction.

Confirming Back and Designing for It

The final safeguard costs one sentence. Before a booking is committed, the resolved slot should be spoken back in terms that cannot be misread: the weekday and the date together, an explicit part of day rather than a bare number, the caller's own language, and the zone if they may be abroad. Repeating the caller's ambiguous words back confirms nothing, so "parso" should become "Thursday the twelfth". This is also where a system should surface what it did not resolve: if the caller said "next week sometime" and Tuesday was chosen, saying so gives them a chance to object.

Everything else is unglamorous engineering. Fix and record one anchor per conversation. Normalise every extracted value with an explicit type and an honest degree of underspecification. Do the arithmetic in code and validate it against real operating hours, working days and a maintained festival calendar. Treat recurrence as a first-class pattern. Build the test set from actual calls rather than invented sentences, including dhaai, paune, bare "baje" times, parso in both directions, and callers abroad.

Above all, confirm. The measure of a scheduling conversation is not whether the system parsed the phrase, but whether the caller and the calendar ended up agreeing on the same moment. At Cirio we treat that agreement as the point of the call, because a booking on the wrong day is indistinguishable, to the customer, from not having called at all.

Frequently Asked Questions

What is temporal expression resolution?

Temporal expression resolution is the task of converting a phrase that refers to time, such as "next Tuesday evening" or "in ten days", into an exact machine-readable value such as a date and time or a duration. It requires an anchor, usually the moment of the conversation, because most spoken time references are relative rather than absolute.

Does "parso" mean the day before yesterday or the day after tomorrow?

In Hindi it can mean either, and the tense of the surrounding sentence usually decides which. In a booking conversation the intent is almost always future, but a system should still confirm the resolved date rather than assume.

How should a voice agent handle "5 baje" without am or pm?

A bare clock time in a 12-hour form is ambiguous, so the system should narrow it using context such as business hours, any part-of-day word the caller used, and whether one of the two candidates is already in the past. If ambiguity remains, it should ask a short clarifying question.

Why does Indian Standard Time simplify scheduling?

India uses a single time zone, UTC+05:30, and does not observe daylight saving time, so there are no missing or repeated local hours to handle. Complexity returns when callers are abroad, since their local reference time and calendar assumptions may differ from the server's.

Put Voice AI to work for your business

Deploy an AI agent that handles calls in Hindi, English, and more in under a minute.

Start free: it's instant →