Go deeper · Ai at work
What else can AI do with language besides chat?
A lot. Long before chatbots, computers translated text, turned speech into captions, read text aloud, sorted reviews by mood and pulled dates and names out of messages. This wider field is called natural language processing. Large language models are one powerful part of it, and smaller task-specific systems still do much of the everyday work.
What you will be able to do
Name five language and speech tasks and their inputs and outputs, explain how speech recognition goes from audio to text and why it makes more errors for some voices, calculate a word error rate, and show where a simple sentiment rule breaks.
Before you start
Helpful but not required: Is every AI a chatbot? for how approaches differ, and how transformers work for tokens and next-word prediction.
1 · You have already seen this
Your phone already listens, reads and talks
Many everyday tools work with words, and most of them never chat with you.
Say a text message out loud and your phone types it for you. A map app reads the next turn aloud in the car, and a screen reader reads web pages aloud for people who are blind or have low vision. When an email mentions “Thursday at 4 pm,” the app may offer to put it in your calendar.
All of these are natural language processing, the part of AI that works with human language, written or spoken. Chatbots are only the newest member of this family. The rest of this page shows how one of these jobs, turning speech into text, works, lets you try a simple rule that judges the mood of a review, and shows where it goes wrong.
The big picture
Language AI is a whole field
Chatbots are the newest and loudest member of a large family of language tools.
Natural language processing (NLP) is the part of AI that works with human language: written text and, together with speech technology, spoken words. It is older than most people think. Researchers built translation programs in the 1950s, and spam filters, spell checkers, search engines and voice dialing all came long before today’s chatbots.
The methods changed over the decades. Early systems used hand-written rules and word lists. Later systems learned statistics from large collections of text. Since the 2010s most use neural networks. The transformer, the design inside today’s chatbots, was first presented in 2017 as a better way to translate between languages.
A large language model (LLM) is a very large model trained on huge amounts of text to predict the next piece. Most widely used chatbot LLMs are big transformers, though other designs exist too. Because it saw so much text, you can ask one model to translate, summarize or judge the mood of a review. That does not make the older tools obsolete. A small model built for one job can be faster, cheaper and easier to test, and it can run on a phone without sending your words to a company’s servers. Many products combine both.
| Task | Input → output | An everyday example | A typical mistake |
|---|---|---|---|
| Machine translation | Text in one language → text in another | Reading a sign or a web page in another language | Word-for-word meaning that misses an idiom, or a guessed gender for “the doctor” |
| Speech recognition | Audio → text | Live captions, voice typing, voice commands | Mishearing names, accents, quiet speakers or noisy rooms |
| Speech synthesis (text-to-speech) | Text → audio | Screen readers, audiobooks, station announcements | Wrong stress or pronunciation; cloned voices can be misused |
| Sentiment analysis | Text → positive, negative or neutral | Sorting thousands of product or app reviews | Sarcasm, mixed opinions, slang |
| Information extraction | Text → structured facts (who, what, when, where) | Turning “Book club meets Thursday at 4 pm in Room 12” into a calendar entry | Mixing up which date belongs to which event |
| Chat and writing help | Text → new text | Chatbots, summaries, drafting | Fluent answers that are wrong (hallucinations) |
Predict first
Your phone shows live captions for a video, even in airplane mode. Is a large chatbot model doing that?
Show the answer
Probably not. Captioning with no internet connection needs a speech recognition model small enough to run on the phone itself. It is a task-specific system: audio in, text out. It cannot chat, and it does not need to.
2 · What happens, step by step
From sound to text
Speech recognition (also called automatic speech recognition) turns audio into written words in three broad steps.
1. Audio. A microphone measures tiny changes in air pressure thousands of times per second. To a computer, a spoken sentence is a long list of numbers.
2. Features. The software cuts the sound into short overlapping slices and measures how loud each range of pitches is in each slice. One well-known model, Whisper, uses slices 25 milliseconds long, starting every 10 milliseconds. The result is a grid of numbers, like a picture of the sound.
3. Model to text. A trained model turns that grid into the most likely words. It learned from many hours of audio paired with correct transcripts, so it also “knows” which word sequences are common. That helps with mumbled words, and it is also why it may write a common word where you said an unusual name.
Errors are not spread evenly. Background noise, people talking at once, accents and dialects the model rarely heard in training, and languages with little recorded audio all raise the error rate. A 2020 study tested five commercial systems on interviews with 115 American speakers and found an average word error rate of 0.35 for Black speakers and 0.19 for white speakers. The Whisper study found that the less training audio a language had, the more errors the model made in it.
Going deeper · step 3
Word error rate: count the fixes
To compare speech recognizers, people count how many word edits it takes to fix the output.
Start with a reference: what was really said, typed carefully by a person. Line it up with the machine’s output. Every wrong word is a substitution (S), every missing word a deletion (D) and every extra word an insertion (I). Then:
WER = (S + D + I) ÷ N, where N is the number of words in the reference.
Lower is better. A WER of 0 means a perfect transcript. One number for a whole test set hides a lot, so good evaluations report WER separately for different groups of speakers, accents, languages and noise levels, just as Good at what, for whom, compared with what? recommends for any model.
Predict first
A recognizer hears someone say “stop” and writes “well stop it now.” What is the word error rate?
Show the answer
3 ÷ 1 = 300%. Nothing was substituted or deleted, but three words were inserted, and the reference has only one word. Because insertions count, WER can go above 100%.
Worked example
Word error rate for one caption
- ReferenceWhat the speaker really said, typed by a person: “please renew the two library books on my card today.” That is 10 words.
- HypothesisWhat the speech recognizer wrote: “please review the library books on my card to day.”
- Line them upplease = please. renew → review is a substitution. the = the. two is missing: a deletion. library, books, on, my, card all match. today → to is a substitution, and the extra day is an insertion.
- CountSubstitutions S = 2, deletions D = 1, insertions I = 1. Errors = 2 + 1 + 1 = 4.
- DivideWER = (S + D + I) ÷ words in the reference = 4 ÷ 10 = 0.40, or 40%.
- Read itFour errors in ten words. Notice that the dropped “two” changes the meaning more than “to day” does, but WER counts every error the same.
The sentence is made up. Tools line the words up automatically by finding the fewest edits that turn one word list into the other; for this pair that minimum is 4, which we checked by computer.
Other language jobs
Translation, voices, moods and facts
Machine translation maps a sentence in one language to a sentence in another. Modern systems translate whole sentences at once, which handles word order much better than word-for-word lookup. They still struggle with idioms, rare languages and words whose meaning depends on context, and they can quietly add assumptions, for example guessing that “the nurse” is a woman when the original language did not say.
Speech synthesis, or text-to-speech, runs the other way: text in, audio out. It powers screen readers for people who are blind or have low vision, and it reads directions aloud in a car. Today’s voices can sound very human, and some systems can copy a real person’s voice from a short recording, which is why voice deepfakes are a real concern. See creative AI for audio generation.
Sentiment analysis labels a text as positive, negative or neutral. A shop can use it to sort thousands of reviews. Information extraction pulls structured facts out of text: names, places, dates, prices. When your email app offers to add “Thursday at 4 pm” to your calendar, that is information extraction. Both can be done with a big general model or with a small model trained only for that job.
The simplest sentiment method is a word list, also called a lexicon: count the happy words, subtract the unhappy ones. In 2002 researchers showed that machine learning trained on labeled movie reviews beat word lists chosen by people. The lab shows you why.
3 · Try it
Break a simple sentiment rule
Eight made-up library reviews, one word-list scorer. Switch the negation rule on and off and look for the failures.
Predict first
With no negation rule, what score does “The new graphic novel was not bad at all.” get?
Show the answer
−1, negative. The only list word is “bad,” so the rule says negative, even though the writer liked the book. A negation rule that flips “bad” after “not” fixes this one. Turn it on in the lab and see what it breaks instead.
Lab
A word-list sentiment scorer
Turn the negation rule on and off. For each review, compare the rule’s verdict with what the writer meant. Then type your own sentences and try to fool it.
A simple rule, not a real AI. Runs in your browser on eight made-up reviews; nothing is sent anywhere.
Green words add +1, red words add −1, crossed-out words were flipped by a nearby “not,” “no” or “never.” The verdict is positive above 0, negative below 0, and neutral or mixed at exactly 0.
The interactive version needs JavaScript. Here is the same lab as a table. The rule adds +1 for each word on the positive list and −1 for each word on the negative list. With the negation rule on, a word such as “not” flips the next list word within three words. Without negation the rule agrees with the writer on 2 of 8 reviews; with it, on 5 of 8. Look at the sarcastic printer review and “Not only fun but also helpful.”
Positive list: good, great, love, loved, helpful, friendly, quiet, fun, easy, best, nice, enjoyed. Negative list: bad, boring, slow, noisy, broken, worst, hate, rude, confusing, complaints, complaint, late. Negation words: not, no, never, isn’t, wasn’t, don’t, doesn’t, didn’t, can’t, won’t, aren’t, weren’t.
| Review (made up) | What the writer meant | Negation rule off | Negation rule on |
|---|---|---|---|
| The study room was quiet and the staff were friendly. | positive | +2 → positive ✓ quiet (+1), friendly (+1) | +2 → positive ✓ quiet (+1), friendly (+1) |
| The new graphic novel was not bad at all. | positive | -1 → negative ✗ bad (−1) | +1 → positive ✓ bad (flipped to +1) |
| Oh great, the printer is broken again. Just great. | negative | +1 → positive ✗ great (+1), broken (−1), great (+1) | +1 → positive ✗ great (+1), broken (−1), great (+1) |
| I did not love the ending, but the art was great. | mixed | +2 → positive ✗ love (+1), great (+1) | 0 → neutral or mixed ✓ love (flipped to −1), great (+1) |
| The self-checkout is not easy to use and often slow. | negative | 0 → neutral or mixed ✗ easy (+1), slow (−1) | -2 → negative ✓ easy (flipped to −1), slow (−1) |
| No complaints: the kids loved the reading club. | positive | 0 → neutral or mixed ✗ complaints (−1), loved (+1) | +2 → positive ✓ complaints (flipped to +1), loved (+1) |
| Not only fun but also helpful. | positive | +2 → positive ✓ fun (+1), helpful (+1) | 0 → neutral or mixed ✗ fun (flipped to −1), helpful (+1) |
| The plot dragged on forever. | negative | 0 → neutral or mixed ✗ no list words | 0 → neutral or mixed ✗ no list words |
What the lab shows. Without negation the rule agrees with the writer on 2 of 8 reviews; with negation, on 5 of 8. It still misses sarcasm (“Oh great, the printer is broken again”), it flips “Not only fun” by mistake, and it knows nothing about words missing from its lists, such as “dragged.” Every patch creates new edge cases. That is why most sentiment tools today learn from many labeled examples instead, and why even they still stumble on sarcasm.
4 · Common mix-ups
Language tools are easy to misjudge
Four ideas that trip people up, and what to think instead.
Mix-up: Language AI means chatbots.
Better: Chatbots are one product. Translation, captions, screen readers, spam filters and review sorting are language AI too, and many run on small task-specific models.
Mix-up: If a large language model can do a task, nobody needs a specialized system for it.
Better: Specialized systems can be faster, cheaper, work offline, keep data on your device and be easier to test. Which one fits depends on the job.
Mix-up: A speech recognizer that works well for me works well for everyone.
Better: Accuracy depends on accent, dialect, language, noise and microphone. Ask for error rates by group.
Mix-up: Sentiment analysis understands feelings.
Better: It predicts a label from patterns in text. Sarcasm, mixed opinions and slang often fool it.
5 · Check yourself
Three quick questions
Answer in your head or on paper before you open each one.
The reference is “meet me at the north gate” (6 words). The recognizer writes “meet me at the gate now.” What is the WER?
Check your answer
“north” is missing (1 deletion) and “now” was added (1 insertion). WER = (0 + 1 + 1) ÷ 6 = 2 ÷ 6 ≈ 0.33, or about 33%.
A word-list scorer calls “I can’t say I loved it” positive. Explain the mistake and one reason a fix is hard.
Check your answer
It counts “loved” as +1 and ignores the meaning of “can’t say.” A negation rule would flip “loved” here, but the same rule wrongly flips other sentences, such as “Not only fun but also helpful.” Language has too many patterns to cover with hand-written rules, which is why tools learn from labeled examples.
A school wants automatic captions for assemblies. Name two things to check before relying on them.
Check your answer
Good answers include: test the WER on recordings from your own hall, with your own speakers; check accuracy for different accents and for names; check whether the audio is sent to a company’s servers; keep a person who can fix important errors.
In the real world
Helpful, with limits
Application
Captions for everyone
Automatic captions help people who are deaf or hard of hearing, people in noisy places and language learners. For lectures or meetings, a human check of names and key terms still matters.
Limitation
Uneven accuracy
A system can work well for some voices and much worse for others. Ask whose speech was in the training data and whether error rates were reported for different groups.
Connection
Language tools that use tools
A voice assistant chains speech recognition, a language model and other software through APIs and tool calls. An error in the first step travels down the chain.
For younger learners
A grades 3–5 version
No formulas: listen, write, compare.
20 minutes · teacher-led
Be the caption machine
- The teacher reads five short sentences, one with a made-up name and one with the radio playing.
- Students write down exactly what they heard.
- Compare with the real sentences and count the wrong, missing and extra words.
- Ask: which sentences were hardest, and why?
Exit question
Happy or not?
Write “This lunch is not bad” and “Great, it’s raining again” on the board. Would a computer that only counts happy and sad words get them right? Good answers explain that the words say one thing and the writer means another.
Honest limits
What this page leaves out
What this page leaves out
Grammar parsing, question answering, summarization, sign languages, speaker identification and how tokenizers split text are all left out. So is the math inside speech models. Whisper is named only as an example with a published paper, not as the best system; no rankings here. The textbook by Jurafsky and Martin in the sources covers all of it.
Where to go next
Keep going
Inside the models
How transformers work
Tokens, embeddings, attention and next-word prediction, step by step.
Systems
How AI connects
APIs, tool calls and connectors: how a voice assistant hands work to other software.
Step back · 01
Is every AI a chatbot?
The whole family of AI approaches, and where language tools fit.
Sources
Where this comes from
- Daniel Jurafsky and James H. Martin, Speech and Language Processing, 3rd edition draft (free online): sentiment and word lists, information extraction, machine translation, speech recognition and word error rate, text-to-speech.
- Bo Pang, Lillian Lee and Shivakumar Vaithyanathan, “Thumbs up? Sentiment Classification using Machine Learning Techniques,” EMNLP 2002: learned classifiers beat human-chosen word lists on movie reviews.
- Allison Koenecke and others, “Racial disparities in automated speech recognition,” PNAS 117(14), 2020: five commercial systems, average WER 0.35 for Black speakers vs 0.19 for white speakers.
- Alec Radford and others, “Robust Speech Recognition via Large-Scale Weak Supervision” (Whisper), 2022: 25 ms feature windows every 10 ms; error rates rise as training audio per language falls.
- Ashish Vaswani and others, “Attention Is All You Need,” 2017: the transformer, first tested on machine translation.
AI words on this page
Natural language processing, Speech recognition, Machine translation, Sentiment analysis, Large language model, Transformer, Token, Training data, Bias, Evaluation, Deepfake, Hallucination
All definitions live in one place: AI words in plain language →
Review status
Written by ARQIV. Last checked October 2026. Independent subject review and classroom pilots are still to come, so treat the activities as drafts and tell us about anything that is wrong or unclear.