¿Qué es el reconocimiento automático del habla (ASR)?
Automatic speech recognition (ASR) turns spoken audio into text. Learn how ASR works on phone calls, why phone audio is hard, and how accuracy is scored.
Respuesta rápida
Automatic speech recognition (ASR) is technology that turns spoken audio into text. A recognizer takes an audio stream, applies a model of speech and language, and returns a transcription. On phone calls it is harder than on clean recordings because of narrowband audio, noise and varied accents.
ASR is the listening half of a voice assistant and the engine behind call transcripts. Everything an AI agent does with a caller's words starts with ASR getting those words right.
What is automatic speech recognition?
ASR converts speech to text. "Speech-to-text" is the everyday name, and the two terms are used for the same thing. The W3C Speech Recognition Grammar Specification describes a speech recognizer as a user agent that takes in one or more grammars, which describe the words and patterns to listen for, plus an audio stream, and produces descriptions of the speech it detected. It notes that most practical recognizers include at least a transcription of the words detected.
How does ASR work on a phone call?
- Audio is captured. The call's audio arrives as a stream, often one channel per speaker.
- The audio is prepared. It is cut into short frames and turned into features the model can read. Silence can be set aside with voice activity detection.
- A model predicts words. It combines how the sound looks with what words are likely to follow each other in the language.
- Text is returned. Often as a running transcript while the speaker is still talking, then finalised at the end of an utterance.
- Software acts on the text. It may be stored as a transcript, searched, summarised, or passed to an AI agent that decides what to say next.
Why is phone audio harder for ASR?
- Narrowband audio. Telephone audio carries less detail than studio sound, so similar sounds are easier to confuse.
- Noise. Traffic, wind, and background voices on a mobile call.
- Accents and dialects. A model trained on one variety may stumble on another. See the post on Arabic speech recognition across dialects.
- Cambio de código. Speakers who mix two languages in one sentence.
- Numbers and names. Short, unpredictable words with no context to help.
How is ASR accuracy measured?
The usual score is word error rate (WER): the share of words that were substituted, deleted or inserted compared with a correct reference transcript. NIST publishes the Speech Recognition Scoring Toolkit (SCTK), which includes the SCLITE scoring tool, for evaluating recognizer output. The WER entry y la guía de measuring call transcription accuracy show how to score your own calls. A WER from someone else's test set says little about your callers, so measure on your own audio.
How does ASR differ from text-to-speech and voice activity detection?
| ASR | Text-to-speech | Voice activity detection | |
|---|---|---|---|
| Input | Speech audio | Text | Audio |
| Output | Text | Synthetic speech | A decision: speech or not speech |
| Role in a voice agent | Hears the caller | Speaks to the caller | Tells when the caller starts and stops |
| Más información | This page | Text-to-speech | Voice activity detection |
Where does DialerBee use ASR?
Two places on the site's feature pages. Transcripción de llamadas converts calls into searchable text, with the interface, answering machine detection and transcription covering 11 languages. The agente de voz de IA holds the call itself, and leaves behind a recording, a transcript and an outcome. The voice agent has run on live calls in Arabic and English. For other languages, test it on a demo first.
How can you evaluate ASR for your own calls?
- Collect a sample of real calls that covers your languages, accents and noise levels.
- Have people write correct transcripts for them.
- Run the recognizer and score its output against those transcripts as word error rate.
- Look at where it fails: names, numbers, one dialect, noisy lines.
- Repeat on fresh calls from time to time, as your callers and lines change.
Preguntas frecuentes
What does ASR stand for?
ASR stands for automatic speech recognition, technology that converts spoken audio into text.
Is ASR the same as speech-to-text?
In everyday use, yes. Both describe turning speech into text. ASR is the technical term.
How accurate is speech recognition on phone calls?
It varies with the language, accent, audio quality and noise. Phone audio is usually harder than clean recordings, so measure word error rate on your own calls rather than relying on a general figure.
Why does ASR struggle with names and numbers?
They are short, and there is little surrounding context to help the model guess. They also vary widely between callers.
Does ASR work in a call centre?
Yes. It is used to transcribe calls for search and review, and to let an AI voice agent understand what a caller says.
Fuentes
- W3C Speech Recognition Grammar Specification Version 1.0, Recommendation, 16 March 2004 (fuente, consultada el 2026-11-19)
- NIST Information Technology Laboratory, evaluation tools (Speech Recognition Scoring Toolkit, SCTK) (fuente, consultada el 2026-11-19)
- NIST SCTK source repository (fuente, consultada el 2026-11-19)
Lecturas relacionadas: Transcripción de llamadas · Word error rate · Text-to-speech · Voice activity detection · Measuring transcription accuracy · Arabic speech recognition and dialects · agente de voz de IA
Lecturas relacionadas: What Is a Large Language Model (LLM)? · What Is Prompt Injection?
Términos relacionados
¿Qué es el tiempo medio de gestión (AHT)?
El tiempo medio de gestión (AHT) es el tiempo total promedio que un agente dedica a una interacción: conversación, espera y trabajo posterior a la llamada.
¿Qué es la velocidad media de respuesta (ASA)?
ASA es el tiempo que un contacto espera en cola antes de ser atendido. Aprenda la fórmula, su diferencia con el nivel de servicio y por qué puede engañar.
What Is Barge-In in Voice AI and IVR?
Barge-in lets a caller cut into an IVR prompt or AI voice agent and be heard. See how it works, when to turn it off, and how it differs from supervisor barge.
¿Qué son Barge, Whisper y Listen?
Escuchar, susurrar e intervenir son los tres niveles de monitorización en vivo de un supervisor. Qué hace cada uno, cuándo usarlo y qué ven los agentes.
¿Qué es BYOC (Bring Your Own Carrier)?
BYOC (traiga su propio operador) permite conectar sus troncales SIP y operador a una plataforma de comunicaciones, sin usar telefonía integrada.
¿Qué es la mezcla de llamadas (call blending)?
Call blending lets the same agents take inbound calls and make outbound calls from one desktop. See how it works, when it helps, and what to watch.
¿Listo para ver DialerBee en acción?
Solicite una demostración en vivo de 15 minutos o comience una prueba gratuita de 14 días: sin presentaciones, sin compromiso. Las llamadas empiezan en cuanto su propia troncal SIP o PBX está conectada y aprobada.
Prueba gratuita de 14 días · crédito de IA gratuito · sin tarjeta · 11 idiomas · BYOC · controles que apoyan el cumplimiento