Amazon Transcribe

Amazon Transcribe is an Amazon Web Services service that converts audio speech into text using a deep learning process called automatic speech recognition (ASR).

AI Practitioner focus

• Transcribe is managed speech-to-text for batch files and real-time streams, with timestamps/confidence plus options such as speaker/channel identification and PII redaction.
• It does not synthesize speech (Amazon Polly) or translate the resulting text (Amazon Translate), though the services can be chained.

Key points

• Uses automatic speech recognition (ASR), a deep learning process, to convert spoken audio into text.
• Amazon Transcribe Medical converts medical speech to text for clinical documentation.
• Best suited for customer service calls, live broadcasts, and media subtitling.
• Can return timestamps and confidence information and supports batch and streaming transcription; accuracy depends on audio quality, language, vocabulary, and configuration.
• Custom vocabulary — a list or table of domain-specific terms, brand names, acronyms, and proper nouns that Transcribe isn’t rendering correctly; used to bias recognition toward the correct spelling (up to 100 vocabularies per account).
• Custom language model (CLM) — trained on up to 2 GB of domain-related text (e.g., in-domain transcripts, technical reports) so Transcribe learns the context a term is used in, not just its spelling/pronunciation; a request can’t combine a custom language model with automatic language identification.
• Automatic language identification — Transcribe can detect the dominant spoken language(s) in media without a language code being specified; batch jobs can identify multiple languages within a single file, while streaming can identify one language per channel (max two channels) and needs at least two candidate languages supplied up front for faster detection.
• Call Analytics — a Transcribe mode built specifically for two-channel (agent + customer) contact-center audio; on top of the base transcript it adds sentiment analysis, call characteristics (talk time, interruptions, loudness), custom category/keyword matching, issue detection, PII redaction, and generative call summarization, for both post-call (S3 batch) and real-time (streaming) use.
• Subtitle output (SRT/VTT) — batch transcription jobs can generate WebVTT (.vtt) and/or SubRip (.srt) subtitle files directly alongside the transcript, supporting the media-subtitling use case above without manually converting timestamps.

Sources

• What is Amazon Transcribe?
• Custom vocabularies
• Custom language models
• Identifying the dominant languages in your media
• Analyzing call center audio with Call Analytics
• Creating video subtitles (SRT/VTT)

Pricing

• Charges are applied based on the seconds of speech converted per month.