Saltar al contenido principal

Reconocimiento de voz

Definición

El reconocimiento de voz (ASR) transcribe audio a texto. Related areas include speaker identification, speech synthesis (TTS), and spoken language understanding.

Sirve de puente entre multimodal (audio as one modality) and NLP (output is text). Modern ASR is mostly end-to-end neural; self-supervised pretraining (por ej. wav2vec 2.0) reduces the need for huge labeled datasets. Deployed in voice assistants, captions, and meeting tools.

Cómo funciona

Audio (waveform or mel spectrogram) is converted to features (por ej. filter banks, learned representations). An acoustic model (por ej. conformer, wav2vec 2.0 encoder) maps features to frame- or segment-level representations. A decoder (CTC, RNN-T, or attention-based) produce text (characters or subwords). Modern systems are often end-to-end (waveform or features → text in one model). Self-supervised pretraining on unlabeled audio (por ej. wav2vec) then fine-tuning on labeled ASR data improves robustness and reduces labeled data needs.

Casos de uso

Speech technologies apply when the input or output is audio: transcription, assistants, and speaker or synthesis systems.

  • Automatic speech recognition (ASR) for transcription and captions
  • Voice assistants and spoken dialogue systems
  • Speaker identification and speech synthesis (TTS)

Documentación externa

Ver también