Amazon Transcribe
Speech-to-text for audio files and live streams.
❓ What is it?
An automatic speech recognition (ASR) service converting audio to text in batch or real time, with speaker diarisation, custom vocabularies, automatic language identification, PII redaction, and toxicity detection; a medical variant handles clinical dictation.
💡 Why does it exist?
Voice is the highest-bandwidth human channel and the least searchable. Transcription unlocks call analytics, captions, and compliance archives without training an ASR model.
⏱️ When should you use it?
Use it for contact-centre call transcription, meeting notes, video subtitling, and as the first stage of any voice pipeline (voice → text → NLP/LLM).
🗺️ Where does it fit?
Between audio sources and text consumers: batch jobs read recordings from S3; streaming transcription feeds live text to Lambda or Kinesis; Contact Lens uses it inside Amazon Connect.
🔌 How do you integrate it?
StartTranscriptionJob for files in S3 (JSON transcript out), or open a streaming session over WebSocket/HTTP2 for live audio; add custom vocabulary for domain jargon and enable redaction where needed.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Direction matters: speech→text = Transcribe; text→speech = Polly. Trivial-sounding, frequently tested.
- Custom vocabulary is the fix when a scenario complains about brand names or jargon being mis-transcribed.