Back to Blog
NLP9 min read1,885 words

Inverse Text Normalisation: Turning Spoken Words Into Written Records

A recogniser hears "double nine eight four triple one". A CRM field needs 9984111.

Abstract cover of flowing teal sound waves condensing into rows of crisp written digits

A speech recogniser produces words. That is what it is trained to do, and for most purposes it is the right output. But the moment recognised speech is destined for a database rather than a human reader, words are the wrong currency. A CRM contact field does not want "double nine eight four triple one two three five six". It wants 9984111256. An invoice field does not want "two lakh fifty thousand rupees". It wants a number and a currency.

The layer that bridges this gap is inverse text normalisation, commonly abbreviated ITN. It is the mirror image of the normalisation that a speech synthesiser performs before speaking: where TTS normalisation expands written forms into speakable words, ITN contracts recognised words back into written forms. It is rarely discussed, it is almost never evaluated properly, and it is responsible for a surprising share of the data quality problems that show up downstream in Indian voice deployments.

Why raw recogniser output is unusable

Consider a caller reading out a mobile number. Depending on speaker and region, the recogniser might produce any of:

  • "nine eight seven six five four three two one zero"
  • "double nine eight four triple one two three five six"
  • "nau aath saat chhah paanch chaar teen do ek zero"
  • "ninety eight seven six, five four three two one zero"

All four are faithful transcriptions. None of them can be stored, dialled, deduplicated or matched against an existing record. The same applies to amounts, dates, order references and email addresses. Every one of these is a field with a defined written format, and the recogniser has no obligation to produce it.

A system that appears to have understood the caller perfectly can therefore still write garbage into the record. The failure is invisible in transcripts, which read fine, and surfaces only when someone tries to call the number back.

The semiotic classes

Following the same taxonomy used for forward normalisation, the practical work of ITN divides into classes, each with its own conventions and its own failure modes.

  • Telephone numbers. Digit sequences with a defined length, possibly a country code, and no separators in storage. Speakers routinely group, repeat and compress them.
  • Amounts of money. "Fifteen hundred rupees", "ek hazaar paanch sau", "two point five lakh", "saadhe teen lakh". All of these denote a value that must end up as a number, and the lakh and crore forms have no English-language equivalent in most ITN grammars.
  • Dates and times. "Fifth October", "paanch tareekh", "agle mangalwar", "saadhe teen baje". Relative expressions need a reference date, which means ITN cannot always be a pure string transformation.
  • Addresses and PIN codes. Six-digit PIN codes, building and flat numbers, sector and phase numbers. These interleave digits and words in ways that resist simple patterns.
  • Percentages, units and measures. "Eight point five percent", "do sau grams", "twelve hundred square feet".
  • Email addresses and spelled strings. Perhaps the hardest class, discussed below.
  • Ordinals. "Third floor", "teesri manzil", "second Saturday".
  • Identifiers. Order numbers, policy numbers, vehicle registrations. Mixed alphanumeric, no redundancy, and therefore no tolerance for error.

Each class needs its own detection logic, its own conversion, and its own notion of what a valid output looks like.

What makes Indian speech harder

Several difficulties are specific to, or much more acute in, Indian deployments.

Mixed-language numbers in one utterance

A single spoken number may switch languages midway. "Mera number hai nine eight two, do sau teen, five six seven eight" is not unusual. An ITN grammar written for one language will parse part of this and drop or mangle the rest. Worse, it may parse the Hindi portion as a quantity and the English portion as digits, producing an internally inconsistent result.

The same applies to amounts. "Three lakh pachattar hazaar" mixes an English numeral with a Hindi one inside a single Indian-system quantity.

The lakh and crore system

"Two lakh fifty thousand" has to become 250000, and if the output is displayed to an Indian user it should be grouped as 2,50,000 rather than 250,000. Fractional forms are common in speech and have no literal digit reading: dhai lakh is 250000, sawa crore is 12500000, dedh is one and a half, paune is a quarter less. A grammar that handles only "two point five lakh" will silently fail on the forms people actually use.

Double, triple and digit compression

Indian English speakers frequently compress repeated digits: "double five" for 55, "triple one" for 111, occasionally "double double two" for 2222. This is a productive convention, not a fixed vocabulary, and it can appear anywhere inside a longer sequence. It also interacts badly with the word "double" in other senses, as in "double bed" or "double payment".

Spelled-out strings and the phonetic alphabet

When a caller spells an email address or a reference code, they rarely spell cleanly. Indian callers commonly disambiguate letters with place names and common words: "B for Bombay", "M for Madras", "S for Sam", "D for Delhi". This is an informal alphabet, not the NATO one, and it varies by speaker.

The ITN task here is to recover B, M, S, D and discard the disambiguating word. That requires distinguishing "B for Bombay" as a spelling gesture from a genuine phrase. Email addresses compound the problem, since the recogniser will produce "at the rate", "at", "dot", "underscore" and "dash" as words, and any of these can also occur as ordinary content. "rajesh dot kumar at the rate gmail dot com" needs to become rajesh.kumar@gmail.com, including the lowercasing and the removal of the spaces the recogniser inserted between letters.

Punctuation, capitalisation and truecasing

Most recognisers output unpunctuated, uncased text. Restoring sentence boundaries, commas and capitalisation is a closely related task, usually handled by a sequence labelling model over the recognised tokens.

For record keeping this matters less than the numeric classes, with two exceptions. Proper nouns need capitalisation if names go into a contact record, and sentence segmentation affects anything downstream that reasons over the transcript. A misplaced boundary can attach a number to the wrong sentence and therefore to the wrong field.

Ambiguity and the need for context

"Twenty five" is a quantity, an age, a temperature, a price, part of a date, or two digits of a longer identifier. The words are identical in every case. Resolution depends on context that may be several turns away.

  • Local context helps: a preceding "rupees" or a following "years old" is usually decisive.
  • Field context helps more: if the system is currently collecting a PIN code, digit interpretation is almost always correct, and a six-digit constraint is available.
  • Global context is sometimes necessary: whether "five six" continues a phone number depends on how many digits have already been collected.

This is why ITN that runs as a blind post-process over a finished transcript performs worse than ITN that knows what is being asked for. The information required to disambiguate is the information the dialogue already has.

Rules, neural models, and the hallucination risk

Two broad approaches dominate.

Weighted finite-state transducers. ITN grammars compiled into WFSTs are the traditional approach and remain widely used; the open-source NeMo text processing toolkit is a well-known example of this style. Their properties are exactly what high-stakes fields need. The transformation is deterministic, auditable, and constrained to outputs the grammar permits. A grammar cannot invent a digit that was not in the input. The cost is labour: each class, each language and each convention must be written and tested, and Indian conventions are not covered by anything off the shelf.

Neural sequence models. Treating ITN as a sequence-to-sequence problem handles disfluency, hesitation, self-correction and unanticipated phrasing far more gracefully. "My number is nine eight, sorry, nine seven double six..." is exactly the kind of input a grammar handles badly and a learned model handles well.

The risk is well documented in the forward direction and applies equally here: a fluent model can produce a plausible output that is not what was said. A transducer that fails, fails visibly. A neural model that fails produces a well-formed ten-digit number that belongs to somebody else. For a phone number, an amount or a policy reference, that is the worst available failure mode, because nothing downstream will flag it.

The practical resolution most teams reach is hybrid. Learned components handle detection, segmentation and messy input; deterministic conversion and explicit validation guard the fields where a wrong value is expensive.

Validation is not optional

Whatever produces the value, high-stakes fields should be checked before they are written.

  • Length and format constraints. Indian mobile numbers are ten digits and begin with 6, 7, 8 or 9. PIN codes are six digits. These checks catch a large fraction of compression and drop errors at essentially no cost.
  • Checksums where they exist. Some identifier schemes carry check digits. Where they do, verifying them is free accuracy.
  • Plausibility ranges. An amount three orders of magnitude outside the expected range for a transaction type is more likely to be a lakh or crore parsing error than a real value.
  • Round-trip confirmation. For the most critical fields, reading the written value back to the caller in spoken form and asking for confirmation is the only check that verifies the whole chain, recogniser and ITN together.

The failure to validate is what turns an ITN error into a business problem. A wrong digit written silently into a record propagates into callbacks that never connect, messages delivered to strangers, duplicate contacts that defeat deduplication, and reconciliation work that nobody budgeted for. The cost is incurred far from where the error occurred, which is why it is so often misattributed to the recogniser.

Evaluating ITN on its own terms

Word error rate will not tell you any of this. WER is conventionally computed over spoken-form text, with normalisation removed from both hypothesis and reference precisely so that formatting differences do not pollute the measurement. A system can post an excellent WER and write unusable records.

ITN needs separate evaluation:

  • Per-class accuracy. Measure phone numbers, amounts, dates and identifiers separately. Aggregate numbers hide the classes that matter most, because they are usually the rarest tokens.
  • Exact match at the field level. For a phone number, partial credit is meaningless. The field is right or it is not.
  • A real test set. Build it from actual call audio in the target languages, including code-mixed utterances, compressions, self-corrections and spelling sequences. Synthetic test cases generated from a grammar will only test what the grammar already knows.
  • Error type analysis. Distinguish recognition errors from ITN errors. If the recogniser heard the wrong digit, a better grammar will not help.

ITN is an unglamorous layer that sits at the boundary between speech and structured data, and it is usually the last thing anyone instruments. For any voice system whose purpose is to leave behind a record that someone will act on, it deserves the same scrutiny as the recogniser in front of it. At Cirio, we regard the quality of what ends up in the record as inseparable from the quality of the conversation that produced it.

Frequently Asked Questions

What is inverse text normalisation?

Inverse text normalisation, usually abbreviated ITN, converts the spoken-form words produced by a speech recogniser into conventional written forms. It turns "nine eight seven six five" into 98765 and "two lakh fifty thousand rupees" into an amount such as ₹2,50,000. It is the reverse of the normalisation step that text to speech performs before synthesis.

Why is inverse text normalisation harder for Indian languages?

Indian speech frequently mixes languages within a single utterance, so a number may begin in one language and finish in another. The Indian numbering system adds lakh and crore, and speakers also compress digits using conventions such as double and triple. Rules and training data developed for English do not cover any of this.

Does inverse text normalisation affect word error rate?

Not directly, because word error rate is usually computed on normalised spoken-form text with the ITN step removed. A system can have excellent word error rate and still produce written output that is unusable in a database. ITN therefore needs its own evaluation, ideally measured per field and per semiotic class.

Should inverse text normalisation use rules or a neural model?

Rule-based transducers are predictable and auditable, which matters for fields such as phone numbers and amounts where a plausible but wrong value is dangerous. Neural models handle disfluency, context and unusual phrasings more gracefully but can produce fluent errors. Many practical systems combine the two and apply validation checks to high-stakes fields regardless of which approach produced the value.

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 →