Streaming against batch transcription
In one sentence
Streaming transcription emits words as they are spoken and revises them as more audio arrives, while batch transcription waits for the whole recording and returns a more accurate finished transcript.
Definition
Batch transcription takes a finished recording and returns a complete transcript. Streaming transcription writes words down as they are spoken and revises them as more audio arrives.
Conversation needs streaming, because you cannot wait for someone to stop talking before you start thinking. Accuracy prefers batch, because seeing the whole sentence resolves the ambiguous words in it.
This is a genuine architectural tradeoff rather than a better-or-worse choice, and understanding it explains several otherwise puzzling behaviors in voice agents.
Batch transcription
- The model sees the entire utterance, including what came after any ambiguous word.
- Later context disambiguates earlier words, which is why batch accuracy is higher.
- Suited to recordings, meeting notes, compliance archives and analytics.
- Unusable for conversation, because it cannot start until the speaker has finished.
Streaming transcription
- Emits partial results as audio arrives, typically within a few hundred milliseconds.
- Revises earlier guesses as more context accumulates. A word can change after it has already been emitted.
- Produces two output types: interim results, which are provisional and will change, and final results, which are committed.
- Accuracy is lower than batch on the same audio, because the model is guessing with less context.
The interim results trap
- Building product logic on interim results is a common and painful mistake. They flicker, change and disappear.
- A system that triggers an action on an interim result may act on a word that was subsequently revised.
- The correct pattern is to display interim results if you are showing live transcription, and to act only on finals.
Where the two combine
- Many production systems use streaming for the conversational path and batch for the stored transcript, so the visitor gets responsiveness and the archive gets accuracy.
- This is more expensive, since the audio is transcribed twice, and it is often the right decision where the transcript has compliance or analytics value.
- It also means the transcript in the dashboard may not exactly match what the agent acted on in the moment, which is worth knowing when debugging.
Latency contribution
- Streaming transcription finalization is a component of time to first audio. The system must commit the final words of the utterance before the model can respond.
Common misconception
That streaming is simply the real-time version of the same thing. It is a different operating mode with measurably lower accuracy, and quoted accuracy figures are frequently batch figures.
Why it matters commercially
Published word error rates are often measured in batch mode on clean audio, and a conversational deployment operates in streaming mode on real audio. Both differences push actual accuracy below the published number.
In voice specifically
A text chat has no interim stage. The visitor presses enter and the words are final, so there is no revision to guard against and no accuracy penalty for replying quickly. Speech forces the choice: respond on words that might still change, or wait until they cannot and lose the flow of the conversation.
Where AsqVox fits
Answering a visitor in real time requires streaming. The transcript that lands in the dashboard can be produced by a separate batch pass, which is how a deployment gets both a responsive conversation and a clean archive.
Visual
Guessing early, or knowing late
Published accuracy is usually batch. Your product runs streaming.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Whisper Large-v3 records around 2.7 percent word error rate on the clean LibriSpeech test-clean benchmark and roughly 8 to 12 percent on real-world English audio.
2.7% clean, 8 to 12% real worldindependentWhisper Large-v3 on LibriSpeech test-clean, against reported real-world English performance, 2026 - These are usually batch figures on clean audio. A streaming conversational deployment on real audio operates below them on both counts.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate under production conditions.
5.26% to 6.84%vendor claimDeepgram Nova-3 reported figures, 2026 - Published by the provider. Confirm whether a quoted figure is streaming or batch before comparing it to anything.
Published accuracy figures are typically measured in batch mode on clean audio, so a streaming conversational deployment on real audio operates below them on both counts.
-industry rangeSpeech recognition benchmarking convention, 2026 - Two separate gaps stack: batch against streaming, and clean audio against real audio. Both push actual accuracy under the published number.
Streaming transcription finalization is a component of the time to first audio budget, which runs against a sub-second target.
Part of a sub-second budgetindustry rangeLatency budgeting convention, 2026 - The system has to commit the final words of the turn before the model can respond, so finalization time lands directly on first-word latency.
A 2026 Coval and Gradium benchmark found a speed against accuracy tradeoff across recognition providers, with no single leader on both.
No provider leads on bothindependentCoval and Gradium benchmark, 2026 - The same streaming-versus-batch tension expressed at the provider level. Any accuracy comparison that ignores latency is incomplete.
No published benchmark quantifies the accuracy difference between streaming and batch modes for the same model and the same audio, despite it being directly relevant to every conversational deployment.
-no reliable figureThe difference is real and directional, batch is more accurate, but the size of it for a given model is not something you can look up.
Examples
In practice
A voice agent occasionally acts on a misheard word that appears correctly in the stored transcript, which makes the bug look impossible. The cause is that the action fires on an interim streaming result, while the archived transcript is produced by a separate batch pass. Moving the action triggers to final results only resolves it, and explains the discrepancy between what the agent did and what the transcript says.
The everyday version
Streaming means the system starts writing down what your customer is saying while they are still saying it, so it can reply quickly. It sometimes changes its mind about a word once it hears the rest of the sentence. Batch means waiting for them to finish, which is more accurate and far too slow for a live conversation.
Usage
Who says it
- Speech engineers use the distinction precisely, and it is one of the first architectural decisions in a build.
- It appears in provider documentation as separate API endpoints, one for streaming and one for batch.
Where it turns up
- On a spec sheet it sits next to transcription mode, interim results, latency, accuracy figures and whether the quoted accuracy is batch or streaming.
- Asking which mode a quoted accuracy figure refers to is a good diagnostic question.
Common misuse
- Quoting batch accuracy figures for a streaming deployment.
- Building action triggers on interim results.
- Assuming the archived transcript matches what the agent acted on in a dual-path architecture.
Questions people ask
What is the difference between streaming and batch transcription?
Batch processes a complete recording and returns a finished, more accurate transcript, because it can see the whole utterance and let later words disambiguate earlier ones. Streaming emits words as they are spoken and revises them as more audio arrives, trading accuracy for responsiveness. Conversation needs streaming; a stored archive can use batch.
Why should I not act on interim transcription results?
Interim results are provisional and get revised as more audio arrives, so a word can change after it was emitted. If you trigger an action on an interim result, you may act on a word that no longer exists. The correct pattern is to display interim results if you are showing live text and to act only on final, committed results.
Why is my agent's live transcript different from the stored one?
Many systems run streaming for the conversation and a separate batch pass for the archive. The two are transcribed independently, so the stored transcript can be more accurate and may not match what the agent acted on in the moment. That is expected in a dual-path architecture, and it is worth remembering when a bug looks impossible.
Is a published word error rate measured in streaming or batch mode?
Usually batch, on clean audio. A conversational deployment runs streaming on real audio, and both differences push actual accuracy below the published figure. Ask which mode a quoted number refers to, because the same model scores worse in streaming than in batch on identical audio.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.