How Are Intent and Parameters Extracted from Speech?

• Input utterance
• Intent and parameter fields
• Specified request

What you will learn

• Trace a spoken sentence through text, an intent and parameter fields.
• Extract and normalize a direction, distance or object name.
• Handle missing or ambiguous values by asking instead of guessing.

Before you start

No programming experience is needed; the example uses a hypothetical request in a conversational interface.

What does an intent represent?

An intent is a type of user request in one conversation turn, such as asking for status or requesting movement in the illustrative application. People can phrase one request in different ways, so a designer provides representative training phrases to teach the classifier cues for each intent. These phrases are examples, not a complete list of everything a user might say.

Before that stage, speech recognition converts an audio waveform into text. If the audio becomes “go north for about two hundred metres,” the language-understanding component works on the text and conversation context; it does not directly observe intent in the waveform. A recognition error that changes “north” into another word can therefore affect later understanding, so the system should let an operator inspect the recognised sentence.

How are words separated into fields?

In the illustrative example, “go north for about two hundred metres” maps to intent MoveToDirection, parameter direction = north, and distance = 200 m. The intent says what to do; parameters hold values needed for that action. This is an illustrative design, not a standard schema or a tested recognition result.

An entity is a span of text that carries a value type, such as a direction, distance or target name. The system assigns an entity type so it can interpret “north” as a direction rather than part of a proper name. Normalization converts different expressions into a consistent representation: “two hundred metres” and “200 m” could map to one numeric value and standard unit. The system should not discard a unit or invent one when the sentence leaves it unspecified.

What happens when a value is missing or ambiguous?

Each intent needs to specify which fields are required and what kinds of values are accepted. Dialogflow CX documentation describes filling required parameters and prompting the user for a missing value. In the example, “go north” lacks distance, so the system asks, “What distance would you like?” rather than choosing one itself.

Ambiguity differs from absence. “Go to the dock” may be incomplete if several places are called a dock; “go farther” may not specify a distance. The system should name the uncertain field and ask a focused question. If a value has the wrong type or is unsupported, it should say it cannot handle the request and ask the user to restate it, rather than silently converting it into a similar command.

What does a complete turn look like?

Illustrative design, not a benchmark result.

The user says a phrase containing a direction and distance. Speech recognition returns text; the language-understanding component selects an intent; entity extraction finds the direction and quantity phrase; normalization converts them into fields; a validator checks that required fields are present and types are suitable. If a field is missing, the interface asks a question and attaches the answer to the same conversation. If two intents fit or a target name matches several objects, it asks the user to choose. Only a complete, unambiguous request allowed by the application proceeds to command validation.

This decomposition also makes debugging possible: compare the original sentence, recognised text, selected intent, extracted spans and normalized values. If the text is already wrong, inspect audio and recognition; if the words are correct but intent is wrong, inspect classification examples; if the number is wrong, inspect entity type and normalization.

Common misunderstanding

An intent is a complete command. It identifies the kind of task; parameters still specify the direction, distance or object needed.

Frequently asked questions

Must training phrases cover every possible wording?

No. They are examples that help match phrasing; they should cover realistic variation without listing every possible sentence.

Why normalize a distance?

Normalization maps different spoken or written forms to a field with a consistent value and unit that later checks can use.

When should the system ask again?

When a required field is absent, a value is unsupported, or more than one interpretation remains reasonable after extraction.

Summary

• Speech becomes text; text matches an intent and entities become parameters.
• Normalization keeps values and units consistent for validation.
• Missing fields or ambiguous meaning call for a focused question, not a guess.