Glossary
Inverse text normalization (ITN)
Inverse text normalization (ITN) is the step that converts a speech recognizer's spoken-form output into written form, so that 'two hundred' becomes 200. It changes how the text reads, not which words the recognizer heard.
Updated October 17, 20264 cited facts, each traced to its source
Also called ITN.
Inverse text normalization (ITN) is a post-processing step in speech recognition. A recognizer hears your words and writes them the way they were spoken. ITN then rewrites that spoken form as written form. The NVIDIA NeMo team defines it as conversion of "spoken-domain automatic speech recognition (ASR) output into written-domain text to improve the readability of the ASR output" 1.
What ITN changes
Speech has no digits, capitals or symbols, so a recognizer that writes only what it hears produces text like this:
- "two hundred" becomes 200. Azure AI Speech's display text applies this conversion, for example 200 for two hundred 2.
- "three thirty p m" can become 3:30 p.m.
- "doctor smith" can become "Dr. Smith", so a title keeps the written style of a document.
The same spoken words can have more than one written form, so a good ITN system uses the context around them.
How ITN is built
There are two main approaches, and many production systems combine them.
| Approach | How it works | Trade-off |
|---|---|---|
| Rule-based grammars | Hand-written rules, often weighted finite-state transducers, map spoken patterns to written ones | Predictable and easy to audit, but each language needs someone to write the rules |
| Neural sequence-to-sequence models | A transformer learns the mapping from examples | Learns from data rather than hand-written rules, but its errors are harder to predict |
The NeMo paper explains why rules still dominate many systems. Many state-of-the-art ITN systems use hand-written weighted finite-state transducer grammars, because the task has extremely low tolerance to unrecoverable errors 3. A wrong conversion can change the meaning of a sentence, such as a dose, a price or a date. Neural alternatives exist too. A 2021 paper on neural ITN uses transformer sequence-to-sequence models, and it adds a hybrid framework that pairs the neural model with a finite-state transducer to overcome common recoverable errors in production.
Why it matters for dictation
When you dictate into a document, ITN decides how the text looks on screen:
- Numbers and dates are the most visible change. See how to dictate numbers, dates and symbols for the spoken forms that work when a system leaves a number in words.
- Punctuation and capitals often come from a separate step. Spoken punctuation and automatic punctuation do related work, so check which step produced a given change.
- Errors can look like meaning changes. Read the output before you send it, especially amounts, codes and dates.
ITN also affects how accuracy is measured. A word error rate can depend on whether numbers are written as words or digits. Kacprzak and Fraś conclude that without number normalization, a comparison between two recognizers can be misleading 4. When two accuracy numbers come from different tests, check whether either one applied the same conversion. The guide on how accurate speech recognition is explains how test setups change the result.
Custom words and ITN
ITN works on the words the recognizer already produced, so it cannot fix a name that was misheard. A custom vocabulary helps the recognizer choose the right words first.
Format problem or recognition problem?
When dictated text looks wrong, first check what was said. If the words are right but the form is wrong, such as "twenty five" where you wanted 25, the problem is in formatting, and the fix may be a setting or a spoken command. If the words are wrong, such as "fifteen" heard for "fifty", the problem is in recognition, and formatting cannot repair it. A short test passage with numbers, dates and a name you use often shows which step needs attention.
Related terms
- Automatic speech recognition (ASR): the recognizer whose spoken-form output ITN converts.
- Word error rate (WER): the accuracy measure that normalization choices can change.
- Automatic punctuation: punctuation added to the text, often as a neighboring step.
- Custom vocabulary: word lists that help recognition before formatting.
Guides that cover it
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.
- 2.
- 3.
- 4.
Related terms
- Speech Recognition: What It Is and How Well It WorksThe topic this term belongs to
- Acoustic modelAn acoustic model is the part of a speech recognizer that scores how well a short window of audio matches the sounds of speech. Since about 2012 most acoustic models are deep neural networks, and their training data strongly affects accuracy.
- Automatic speech recognition (ASR)Automatic speech recognition (ASR) is software that converts spoken audio into written words. Modern systems are neural networks trained on large amounts of transcribed speech, and they run in the cloud or on the device.
- ConformerA Conformer is a neural encoder for speech recognition that combines self-attention, which sees the whole input, with convolution, which captures local sound patterns. FastConformer is NVIDIA's faster variant of the same design, used in streaming models.
- Connectionist temporal classification (CTC)Connectionist temporal classification (CTC) is a way to train a neural network to turn audio into text without knowing in advance which moment of sound matches which letter. The network labels every short slice of audio with a character or a blank, and the blanks and repeated labels are then removed to leave the transcript.
- Hidden Markov model (HMM)A hidden Markov model (HMM) is a statistical model in which a sequence of hidden states produces the observations you can measure, and each state changes to the next with set probabilities. In speech recognition, the hidden states are units of sound and the observations are short frames of audio.
- Language model (in speech recognition)In speech recognition, a language model (LM) scores how probable a word sequence is in a language. The recognizer combines that score with the acoustic evidence to choose between words that sound alike, such as homophones.
- LibriSpeechLibriSpeech is a free corpus of about 1,000 hours of read English speech, taken from LibriVox audiobooks, and it is one of the most widely used benchmarks for English speech recognition. Its scores describe clean, read speech, which differs from everyday dictation.
- Mel spectrogramA Mel spectrogram is a grid of sound energy over time, with frequency bands spaced to match human hearing. Many speech recognizers compute its logarithm, the log-Mel spectrogram, and feed it to a neural network instead of the raw waveform.
- Real-time factor (RTF)The real-time factor (RTF) is the time a system takes to process audio divided by the length of that audio. An RTF of 1 or lower means the system keeps up with real-time speech. Some leaderboards report the inverse, RTFx, where higher is faster.
- RNN transducer (RNN-T)An RNN transducer (RNN-T) is a neural speech model that reads audio in order and emits words as it goes, so text can appear while the speaker is still talking. It has an encoder for the sound, a prediction network for the words already written, and a joint network that chooses the next output.
- Speaker diarizationSpeaker diarization is the task of working out who spoke when in a recording, by splitting the audio into turns and grouping turns that come from the same voice. It is separate from transcribing the words, and a transcript gains speaker labels only when diarization is run alongside recognition.
- Voice activity detection (VAD)Voice activity detection (VAD) is a component that decides whether a stretch of audio contains speech. Speech recognizers use it to split audio into turns, skip silence and decide when a speaker has finished.
Pages that use this term
- Speech Recognition: What It Is and How Well It WorksTopic
- How Speech Recognition Works, From Sound to TextGuide
- Word Error Rate (WER): How to Calculate and Read ItGuide
- Open-Source Speech Recognition: The Best ModelsGuide
- Speech Recognition Benchmarks and LeaderboardsGuide
- Why does talk to text censor words?Question
- Language model (in speech recognition)Glossary
- Word error rate (WER)Glossary