Aws Transcribe

Transcribe offers two modes: batch (upload audio to S3, start job, poll/retrieve transcript) and streaming (real-time WebSocket transcription for live audio). For agents processing recorded calls, meetings, or voice messages: use batch transcription with speaker diarization to separate speakers, custom vocabulary to improve accuracy for domain-specific terms, and content redaction to mask PII. Confidence is docs-derived.

AWS SDK (boto3): ts = boto3.client("transcribe"). Batch: ts.start_transcription_job(TranscriptionJobName="job1", Media={MediaFileUri: "s3://..."}, MediaFormat="mp4", LanguageCode="en-US", Settings={SpeakerLabels: True, MaxSpeakerLabels: 5}). Poll: ts.get_transcription_job(TranscriptionJobName="job1") until status COMPLETED. Transcript at TranscriptFileUri (S3 URL). Streaming: WebSocket connection with TranscribeStreamingClient (aws-transcribestreaming-client SDK). Custom vocabulary: ts.create_vocabulary(VocabularyName=..., LanguageCode=..., Phrases=[...])

IAM auth: transcribe:StartTranscriptionJob, transcribe:GetTranscriptionJob, transcribe:ListTranscriptionJobs. S3 access: read input, write output. Managed identity for EC2/Lambda. Streaming: AWS signature v4 WebSocket connection. HTTPS enforced.

Job states: QUEUED → IN_PROGRESS → COMPLETED/FAILED. Poll with exponential backoff. Failure reason in FailureReason field. Supported formats: MP3, MP4, WAV, FLAC, OGG, AMR, WebM. Max file size: 4 GB. Max duration: 4 hours. Transcribe SLA: 99.9%. Accuracy varies by audio quality, accent, and background noise.

docs.aws.amazon.com/transcribe covers batch transcription, streaming API, speaker diarization, custom vocabulary, and medical transcription. Getting started: upload audio to S3, start job via boto3, retrieve transcript in under 10 minutes. Free tier: 60 minutes/month for first 12 months. Community via AWS Forums.