Use real-time transcription when text is needed during the conversation; use batch transcription when recorded audio can be processed after the call. Contact centers often need both: a streaming path for live assistance and a controlled post-call path for finalized transcripts and review. Neither architecture guarantees accurate speaker attribution or business decisions without evaluation.

This guide compares the request paths, gives a latency-budget template and explains speaker diarization, scheduling and operational checks. The examples are illustrative; no call recording or customer performance result is implied.

Architecture 1: real-time speech to text

Authorized call audio
    → capture and supported audio encoding
    → authenticated streaming connection
    → partial hypotheses + finalized segments
    → ordered transcript state with timestamps
    → live agent display or bounded assistance
    → approved retention and post-call reconciliation

Partial text can change as more speech arrives. The application must distinguish provisional text from finalized segments and avoid appending every hypothesis as a new sentence. Preserve segment identifiers or the service’s documented update semantics.

A live assistance feature may use provisional text for suggestions, but consequential updates should follow an appropriate confirmation or review boundary. Do not automatically turn an uncertain recognized phrase into an irreversible customer-record action.

Architecture 2: recorded-audio batch transcription

Approved recording + call metadata
    → protected object storage
    → validated job manifest and transcription settings
    → queued asynchronous transcription
    → completion / failure tracking
    → transcript, timing and supported speaker labels
    → quality checks, review and downstream processing
    → retention and deletion workflow

Azure’s batch transcription overview describes asynchronous processing of recorded audio. A submitted job is not a completed transcript. Track state, correlate outputs with the original recording and handle item-level failures.

Post-call processing can support review, search and summarization without meeting the same live latency budget. It still needs a completion objective and a reliable route for failed or delayed jobs. Verify the current API, region, audio format and supported options before implementation.

Real-time vs batch transcription: selection matrix

Batch is not automatically more accurate, and streaming is not automatically unsuitable for final records. Compare the supported service configurations and actual audio. The architecture choice starts with when the text is needed.

Build a streaming latency budget by stage

Measure from the relevant spoken event to useful display or assistance, not only the provider’s recognition time. Audio buffering, network transfer, endpointing, transcript updates and downstream analysis all contribute.

This worksheet intentionally provides no invented universal millisecond target. Set a user-facing objective, instrument each stage and test under relevant concurrency and network conditions. Report partial-text delay separately from final-segment delay.

A reconnect needs an explicit recovery strategy. Check what audio was acknowledged or retained and whether replay could duplicate transcript content. Do not assume a streaming session can always resume from an arbitrary point without application handling.

Diarization is not identity verification

Speaker diarization groups speech by apparent speaker. A label such as “Speaker 1” does not prove the person’s name or authority. Map agent and customer roles using trusted channel or session metadata where available, and retain uncertainty when attribution is ambiguous.

Azure’s real-time diarization guidance demonstrates speaker-labeled transcription. Its language and diarization configuration guide covers supported configuration choices across transcription modes. Verify the exact feature combination for your API and audio format.

Separate audio channels can provide useful role information when recording topology preserves it; they are not the same as diarizing a mixed channel. Evaluate overlap, interruptions, transfers and multiple participants. A supervisor joining the call can invalidate a simplistic two-speaker assumption.

Batch scheduling, retries and protected recordings

Use a job manifest with stable recording IDs, source locations, language configuration and processing status. Keep a distinct record for submission, completion and downstream publication. Retry transient failures within bounds and avoid reprocessing successful recordings unnecessarily.

Limit access to audio and transcript storage. Avoid long-lived public recording URLs. Apply the organization’s approved notice, access and retention requirements, with responsible review for the operating locations. Do not infer legal compliance from the transcription product’s feature list.

Protect derived data too

Transcripts, summaries, embeddings and search snippets can expose the original conversation. Propagate access restrictions and deletion requirements through downstream systems. Keep sensitive content out of unrestricted operational logs.

If both streaming and batch outputs exist, label their status clearly. A post-call transcript should not silently overwrite reviewed notes without provenance. Link each derivative to its source recording and processing configuration.

Evaluate the business-relevant transcription errors

Use an approved audio set covering relevant accents, languages, channel quality, noise and vocabulary. Obtain reviewed references. Word error rate is useful, but inspect critical entities, quantities, negation, speaker attribution and timestamps separately.

A transcript can have acceptable overall word accuracy while getting an account identifier or the phrase “do not approve” wrong. Evaluate the downstream task as well: a summarizer cannot reliably repair an unsupported transcription guess.

Use the application testing playbook for any LLM-based post-processing. Keep recognition metrics separate from generated-summary quality so the failing stage remains visible.

Common questions

Choose the transcription path around the conversation workflow

For AI post-processing choices, read RAG versus fine-tuning and prompt engineering. To review speech capture, transcription and reliable downstream assistance, discuss your production AI reliability roadmap.