Sentiment analysis

Emerging & Agenticalso: opinion miningalso: sentiment classificationalso: aspect-based sentiment

In one sentence

Sentiment analysis determines whether what someone said was positive, negative or neutral by classifying the language they used, not by trying to read the speaker's emotional state.

Last reviewed 31 July 2026

Not to be confused with Emotion detection and affective computing.

Definition

Sentiment analysis works out whether what someone said was positive, negative or neutral. It reads the content of a conversation rather than trying to read the speaker's emotional state.

That is a smaller claim than emotion detection makes, and a smaller claim is easier to defend.

The distinction from emotion detection is real and worth defending carefully, because the two are marketed interchangeably and stand on very different ground.

What sentiment analysis actually does

  • Classifies text by its expressed valence, usually positive, negative or neutral, sometimes on a graded scale.
  • Works on what was said, which is an observable artifact, rather than on how the speaker feels, which is not.
  • Aspect-based sentiment goes further and attaches sentiment to specific subjects: the delivery was slow but the product was good.

Why it stands on firmer ground than emotion detection

  • The object being classified is the language, not the person. "This is terrible" reads as negative in expression whatever the speaker's internal state.
  • It is a long-established natural language processing task with a substantial research history and well-understood limitations.
  • It makes a smaller claim, and smaller claims are more defensible.

The genuine limitations, which belong on the page

  • Sarcasm and irony are handled poorly. This is a well-known weakness.
  • Negation trips it up. "Not bad at all" is frequently misclassified.
  • It is domain-dependent. Language that reads as negative in one context is neutral in another, and clinical or technical wording is a common source of false negatives.
  • Aggregation hides variance. An average score across a conversation can bury a sharp negative turn at the end, which is usually the part that matters.
  • In voice, transcription error propagates. The sentiment is computed on imperfect text.

How it is actually useful in a voice product

  • Triage. Flagging conversations for review rather than scoring people.
  • Trend detection. Sentiment shifts across a population of conversations surface problems earlier than complaint volume does.
  • Escalation triggers. Detected negative sentiment as one signal among several for handing a conversation to a human.
  • Content diagnosis. Consistently negative sentiment around one topic points at a documentation or product problem rather than a customer problem.

What it should not be used for

  • Individual assessment presented as fact.
  • Any decision about a person made on the score alone.
  • Inferring emotional state, which is the boundary back into contested territory.

Common misconception

That sentiment analysis tells you how customers feel. It tells you how the language they used reads. The two correlate and are not the same, and that gap is exactly the difference between a defensible claim and a contested one.

Why it matters commercially

Sentiment on transcripts is a practical analytics capability with well-understood limits. It is most valuable in aggregate and as a triage signal, and least valuable as a per-conversation verdict. Positioning it accurately earns more trust than overstating it.

In voice specifically

In a voice product the sentiment is computed on a transcript, not on the words as typed, so whatever the speech recognizer got wrong is carried into the classification before it starts. A text chat hands the classifier the exact words; voice hands it an approximation of them.

Where AsqVox fits

The Orb's dashboard surfaces sentiment alongside transcripts. The honest framing is that this is content sentiment used for triage and trend detection, not an assessment of the visitor's emotional state, and that framing holds in the product copy as well as here.

Visual

Sentiment against emotion, a distinction worth defending

Sentiment against emotion, a distinction worth defendingWhat is being classified: the words on the transcript, or the state of the person who said them?Emotion detectionObject of analysis: the person, and a guess at what they feel.The claim: "the person feels angry."Status: an inference about an internal state, scientifically contested.Reads the sound of the voice and reports a named emotion.Infers a person, contestedSentiment analysisObject of analysis: the words, classified by how they read.The claim: "the language reads as negative."Status: an observable artifact, an established NLP task.Reads what was actually said and classifies its valence.Classifies language, defensibleWhere sentiment gets it wrongSarcasm: "brilliant, another outage" scored positive. Negation: "not bad at all" scorednegative. Domain: clinical language scored negative. Aggregation: a conversation thataverages neutral while a sharp negative final turn, the part that mattered, is hidden by theaverage. Used well, it is a triage, trend-detection and escalation signal.Good for triage, not verdicts

A smaller claim, and therefore a defensible one.

The boundary between the panels is the difference between what was said and what was felt. Sentiment classifies the first; emotion detection guesses at the second. In a voice product both run on a transcript, so recognition error reaches them before they start.

Statistics

Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.

There is no single transferable accuracy figure for sentiment analysis. Performance is heavily domain-dependent, and a model that performs well on product reviews may perform poorly on support conversations. Published benchmark figures do not predict performance on a specific corpus.

-no reliable figure

The honest framing for the whole page: an accuracy number quoted without the corpus it was measured on tells you almost nothing about how it will do on yours.

Sarcasm and negation handling remain acknowledged weaknesses across sentiment systems, long-standing and well-documented in the natural language processing literature rather than a vendor-specific limitation.

Known, unsolved weaknessindustry range

Natural language processing literature, 2026 - A documented convention of the field rather than a single measured study. Worth stating because vendors rarely surface it, and it is the failure a buyer will notice first.

In a voice system, sentiment is computed on transcripts carrying roughly 8 to 12 percent real-world word error rate for Whisper Large-v3, or around 5.26 to 6.84 percent in production for Deepgram Nova-3. Transcription error propagates into the sentiment classification.

8 to 12%, 5.26 to 6.84%industry range

Whisper Large-v3 reported real-world English and Deepgram Nova-3 production figures, 2026 - The first band is reported real-world English; the second is a vendor production figure. Carried here to show that the classifier never sees the exact words in a voice product, only an approximation of them.

The EU AI Act prohibits emotion inference in workplace and educational settings from 2 February 2025. Sentiment analysis of content is a different operation, but a system marketed as detecting how employees feel may attract that prohibition regardless of what it is called internally.

Prohibited from 2 Feb 2025industry range

EU AI Act, prohibited practices provisions, 2025 - The label does not determine the classification; the function does. Positioning sentiment as content analysis rather than emotion reading is a compliance point as well as an honesty one.

There is no published benchmark for sentiment analysis accuracy on conversational voice transcripts specifically, as distinct from written text.

-no reliable figure

Written-text benchmarks exist and do not transfer, because spoken transcripts carry disfluency, recognition error and conversational structure that review text does not.

Examples

In practice

A team uses average sentiment per conversation as a quality metric and finds it useless. Switching to sentiment trajectory, the gap between the first and the last third of each conversation, produces a far more useful signal: the conversations that end worse than they began are the ones worth reviewing. The same underlying classification, applied differently, becomes actionable.

The everyday version

Sentiment analysis reads whether the words someone used were positive or negative. It is useful for spotting which conversations went badly and for noticing when a lot of people start complaining about the same thing. What it is not is a reliable read on how any one person was feeling, and it will misread sarcasm every time.

Usage

Who says it

  • Analytics and customer-experience teams, routinely.
  • Product teams, in dashboard and reporting design.
  • Buyers, as an expected analytics feature.

Where it turns up

  • Next to conversation analytics, transcript search, topic detection, escalation triggers and reporting.
  • Worth pinning down whether it is per-conversation, per-turn or aspect-based, because those differ a lot in how useful they are.

Common misuse

  • Presenting it as emotion detection, which imports the scientific and regulatory problems of a different capability.
  • Reporting aggregate sentiment with no trajectory, which hides the conversations that turned.
  • Using a per-conversation score as a verdict on an individual rather than as a triage flag.

Questions people ask

What is the difference between sentiment analysis and emotion detection?

Sentiment analysis classifies the language that was said as positive, negative or neutral, which is an observable artifact and a well-established task. Emotion detection tries to infer the speaker's internal emotional state, usually from the voice, which is scientifically contested. Sentiment makes the smaller and more defensible claim, and the two are marketed interchangeably despite standing on very different ground.

How accurate is sentiment analysis?

There is no single transferable accuracy figure. Performance is heavily domain-dependent, so a model that does well on product reviews can do poorly on support conversations, and published benchmarks do not predict performance on a specific corpus. Sarcasm and negation remain acknowledged weaknesses across sentiment systems, and in a voice product the classification runs on a transcript that already carries recognition error.

Why does sentiment analysis get sarcasm wrong?

Because it classifies the surface language, and sarcasm inverts the meaning of the words while leaving them positive on the surface. "Brilliant, another outage" reads as positive to a classifier looking at word valence. Negation causes the mirror-image error, where "not bad at all" is scored negative. Both are long-standing, well-documented weaknesses rather than a vendor-specific fault.

How should sentiment analysis be used?

As a triage and trend signal, not a verdict. It is most useful for flagging conversations to review, for spotting when many people start complaining about the same topic, and as one escalation signal among several. Report it with trajectory rather than a single average, because an average can hide a sharp negative turn at the end, which is usually the part that matters.

Share this definition

Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.