AI Glossary / Reinforcement learning and robotics / Intermediate

Stt

Definition

stt stands for speech-to-text: the process of converting spoken audio into written text. It is commonly called automatic speech recognition, or ASR. In the supplied PAL Robotics glossary, stt is an alias for asr and refers to transcribing speech captured by the robot's microphone. Transcription is distinct from deciding what the speaker intends or executing a command.

What it means

Speech-to-text turns an acoustic signal into a sequence of written words or other text units. It can support dictation, captions, searchable transcripts, and spoken interfaces. A robot may pass the transcript to a separate language-understanding component that identifies an intended action and its details.
A fluent transcript is not necessarily an accurate transcript. A recognizer can produce plausible words that differ from what was said, especially when audio is noisy, speech overlaps, or terminology is unfamiliar. Downstream software should not confuse linguistic plausibility with evidence that the microphone captured those exact words.

How it works

The system captures audio and converts it into a representation suitable for recognition. A trained model estimates text from the acoustic information, often using patterns learned from speech and language data. Some systems process complete recordings; streaming systems produce results as audio arrives and may revise partial text.
The output may include timestamps or confidence-related information, depending on the implementation. These are additional features, not requirements of the basic speech-to-text definition. Confidence scores also need empirical interpretation; they should not automatically be treated as perfectly calibrated probabilities.
Recognition quality is often measured with word error rate, based on the minimum number of word substitutions, deletions, and insertions needed to align a hypothesis with a reference transcript. Evaluation conventions matter, including how punctuation, numbers, and capitalization are normalized.
For a spoken interface, transcription accuracy is only part of the outcome. The system must also identify when a user is speaking, interpret the request, resolve ambiguous entities, and decide how to handle uncertain or consequential commands. Speech-to-text alone does not provide all of those functions.

Three practical examples

• Robot interaction: transcribe a spoken request before extracting the intended task and destination.
• Accessibility: generate captions that make spoken material available as text.
• Work records: produce searchable draft transcripts of meetings or recorded observations, with appropriate review for the intended use.

Think of it this way

Think of speech-to-text as a listener preparing written notes. It supplies a textual record for another process to use. Reading those notes to decide what to do is a separate step, and a single misheard name or negation can matter more than several harmless wording differences.

Worked example

Consider the fictional ten-word reference command: please bring the blue folder to meeting room seven now. A recognizer returns: please bring the red folder to meeting room seven now. There is one substitution, blue to red, and no insertions or deletions. Word error rate is 1/10 = 10%.
A second recognizer returns: bring the blue folder to meeting room seven. It omits please and now but preserves the other eight words. Its two deletions give word error rate 2/10 = 20%. Under this metric, the first transcript is better. For the task of retrieving the correct folder, however, changing blue to red may be more consequential than omitting those two words. This illustrates why task-specific evaluation should accompany transcript-level metrics.
Now consider a separate test collection containing 1000 reference words. After a consistent minimum-edit alignment, the system has 45 substitutions, 20 deletions, and 15 insertions. Its aggregate word error rate is (45 + 20 + 15)/1000 = 8%. That does not mean 92% of complete commands were perfectly understood; word accuracy and command success are different measurements.
For a tiny reference containing only stop, suppose the hypothesis is stop moving forward. The two inserted words give word error rate 2/1 = 200%. WER can exceed 100% because insertions are counted against the reference length. An empty reference needs separate handling rather than division by zero.
All examples assume lower-case word tokens and the written forms shown. If a system outputs the numeral 7 while the reference uses seven, the evaluation must apply a declared normalization rule. Otherwise formatting differences can change the score without representing a genuine acoustic recognition error.

Illustrative case study

Fictional case: CedarReception adds speech input to a service robot. Quiet-room demonstrations work well, but busy-lobby tests expose errors in room names and numbers. The team evaluates recordings across representative background noise and speaking styles, then checks destination recognition separately from overall word error rate.
For uncertain destination requests, the interaction asks for a targeted clarification before initiating the task. This invented case illustrates interface design around recognition errors; it makes no claim about a specific product's accuracy or default PAL behavior.

Common misconceptions and mistakes

Common mistakes include confusing speech-to-text with text-to-speech, treating transcription as language understanding, and assuming a good average score applies equally to every speaker or environment. Overlapping speech and unusual names deserve separate testing. Partial streaming transcripts may also change, so an application should distinguish provisional text from a final recognition result.

Business use

Businesses should evaluate speech-to-text with representative recordings and the vocabulary of the intended workflow. Useful measures include transcription errors, latency, missed speech, critical-entity accuracy, and completed-task outcomes. The operating process should specify how users correct mistakes and how recordings or transcripts are handled according to the organization's requirements.

People also ask

Is stt the same as ASR?

In this glossary context, yes.

Does it translate between languages?

Transcription and translation are different tasks, though a system can combine them.

Does it tell a robot what to do?

It provides text; interpreting and authorizing an action require additional components.

Check your understanding

Why is the first transcript's WER 10%?

Answer: One of ten reference words is substituted.

Why might the second transcript be more useful for retrieval?

Answer: It preserves the folder color despite omitting two less important words.

Can WER exceed 100%?

Answer: Yes, when insertions are sufficiently numerous relative to the reference length.

Related terms

• Automatic Speech Recognition
• Word error rate
• Speech synthesis
• Speaker diarization
• Acoustic model
• Language model
• transcription
• streaming recognition
• intent recognition
• human-robot interaction

Learn next

Next, compare transcript accuracy with task success on a small set of representative spoken requests. Include names, numbers, negations, background noise, and a documented text-normalization policy.

More in Reinforcement learning and robotics

• Statement - In this robotics knowledge-base context, a Statement is a structured assertion represented as a subject, predicate, and object triple.
• Stochastic multiarmed bandit - A stochastic multiarmed bandit is a sequential decision problem in which a learner repeatedly chooses among actions with uncertain rewards.
• Stochastic policy - A stochastic policy specifies a probability distribution over actions for each relevant state or observation.
• SWARM ROBOTICS - SWARM ROBOTICS studies how groups of robots can produce useful collective behavior through interactions among their members and with the environment.
• Tabular Q-learning - Tabular Q-learning is Q-learning implemented with an explicit table of action-value estimates, one entry for each represented state-action pair.
• Target policy smoothing - Target policy smoothing is a reinforcement-learning technique that perturbs the next action used in a critic's learning target, encouraging value learning...

Sources

PAL Robotics, Glossary, stt and asr: https://docs.pal-robotics.com/edge/glossary Daniel Jurafsky and James H. Martin, Speech and Language Processing, Automatic Speech Recognition chapter, especially evaluation: https://web.stanford.edu/~jurafsky/slp3/16.pdf

Previous: Structuring element Next: Student's t distribution