---
title: "Streaming Speech: Stable Items Within a Partial Segment"
description: "Leviathan Systems — Technical content Published 2026-10-22; updated 2026-10-22."
canonical: "https://www.leviathansystems.co/articles/streaming-asr-item-stability-segment-finality"
last-updated: "2026-10-22"
---

# Streaming Speech: Stable Items Within a Partial Segment

Leviathan Systems — Technical content
Published 2026-10-22; updated 2026-10-22.

With partial-result stabilization explicitly enabled for the selected Amazon Transcribe stream, items marked Stable=true can meet a caption requirement that displayed content must not be revised, even while Result.IsPartial=true. In fictional AS64, those items are blue at E1 and blue kites at E2. A complete S1 export first becomes eligible at E3, when IsPartial=false reports the segment complete. The useful early subset and the complete segment are different deliverables.

## Select the useful subset before the segment is complete

Streaming speech recognition can return a hypothesis whose earlier items are already fixed while other content remains revisable. A caption requirement can use that useful subset without treating the entire hypothesis as finished.

The selected reference is Amazon Transcribe StartStreamTranscription with partial-result stabilization explicitly enabled. AWS's guide describes stable items within partial results. Item.Stable=true identifies an item that will not change in that recognition result. Result.IsPartial=true still describes an incomplete segment; an explicit false value reports a complete segment.

The item property answers which supplied content meets a nonrevision criterion now. The segment property answers whether the selected speech segment is complete. Neither field substitutes for the other. This explanation uses the mutable latest AWS documentation checked October 22, 2026, rather than an assumed immutable service or SDK release.

References: [ASR-GUIDE](https://docs.aws.amazon.com/transcribe/latest/dg/streaming-partial-results.html), [ASR-ITEM](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Item.html), [ASR-RESULT](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Result.html), [ASR-STREAM](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_StartStreamTranscription.html)

## AS64 fixes the caption requirement and the segment associations

AS64 is an original fictional record-interpretation example. It stipulates supported use of the selected operation with stabilization enabled, one channel and one selected alternative. Every transcript, item and flag below is an invented premise. No audio, API call or provider test produced these records.

Requirement R1 asks which initial caption items are explicitly marked unchanging at the current observation, leaving the remaining hypothesis provisional. Requirement R2 asks for the complete selected segment S1. No response-time target is supplied.

The case independently associates E1, E2 and E3 with S1, and E4 with later segment S2. E1 through E4 name successive observations, with no elapsed times assigned. These segment labels are explanatory associations, not invented API fields. The Result reference describes ResultId as a unique result identifier; it does not establish cross-revision constancy here. Equal text or timing is not substituted for the independently supplied associations.

Each listed word has Type=pronunciation. The period at E3 has Type=punctuation. The table supplies the complete selected Items sequences and every Stable value used in the example.

References: [ASR-RESULT](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Result.html), [ASR-ITEM](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Item.html)

AS64: fictional selected transcripts and explicit item flags
| Observation / supplied segment | Result.IsPartial | Selected Transcript | Item Content: Stable | Source |
| --- | --- | --- | --- | --- |
| E1 / S1 | true | blue cars flow | blue: true; cars: false; flow: false | Planning analysis |
| E2 / S1 | true | blue kites float | blue: true; kites: true; float: false | Planning analysis |
| E3 / S1 | false | blue kites float. | blue: true; kites: true; float: true; .: true | Planning analysis |
| E4 / S2 | true | near hills | near: true; hills: false | Planning analysis |
All contents and flags are supplied case premises. All items are pronunciation items except E3's punctuation item. S1/S2 associations are supplied separately; the table specifies no client correlation or update algorithm.

## E1 supports blue; E2 supports blue kites

At E1, blue is explicitly stable. Cars and flow are explicitly not stable. R1 therefore supports the one-item caption subset blue while the other two items remain provisional. Calling every item provisional because S1 is partial would discard the useful positive evidence attached to blue.

At E2, blue and kites are explicitly stable, while float is not. R1 now supports blue kites: two pronunciation items inside the same still-partial S1. R2 continues to wait. Requiring segment completion for this narrower caption choice would postpone content whose nonrevision property is already supplied.

The earlier provisional cars flow has become kites float. These are successive hypotheses for S1, not two completed utterances. The decisive addition is the explicitly stable subset within the current hypothesis. This example does not prescribe how an application stores or updates its caption view.

The marked subset forms a contiguous beginning in AS64. That layout is a case premise, not a universal guarantee about the number or arrangement of stable items.

Float also illustrates the precise meaning of Stable=false. It may change, but it remains float at E3. A false flag requires neither an eventual edit nor a finding that the item is incorrect.

References: [ASR-GUIDE](https://docs.aws.amazon.com/transcribe/latest/dg/streaming-partial-results.html), [ASR-ITEM](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Item.html)

## E3 completes S1; E4 concerns a later partial segment

E3 first supplies IsPartial=false for S1. R2 is now satisfied in the supplied sequence: the selected complete-segment content is blue kites float. Its Items sequence contains three pronunciation items and one punctuation item, giving four items in total. It does not contain four spoken words, four audio frames or four model tokens.

The full stop is an explicit punctuation item in E3, not punctuation assumed from the earlier partial text. Segment-export eligibility follows the supplied completion flag for S1, rather than the presence of a period or a particular text length.

At E4, S2 is partial. Its near item is stable and hills is provisional, so the same R1 criterion supports near for that selected segment. S1 remains complete. A partial result for the later S2 does not reopen S1.

These status decisions concern recognition output. The case supplies neither whole-session completion nor an independently correct transcript. Stability does not assess correspondence with the original speech, and segment completion does not supply an application accuracy result. No ground-truth comparison is available from which to calculate one.

References: [ASR-RESULT](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Result.html), [ASR-ITEM](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Item.html)

## Removing Stable changes what R1 can conclude

Now consider a countercase that removes all Stable fields from E2 while retaining its text and IsPartial=true. R2 still waits because the segment is partial. R1 now lacks the explicit item-stability evidence its criterion requires.

This is an unsupported selection, not a finding that the actual stable-item count is zero. The Item reference makes Stable optional. Its absence does not supply an unseen true or false value, or establish that stabilization was disabled. Repeated wording across observations likewise does not reveal an omitted flag.

The reverse substitution also fails. Even if every currently listed item were described as stable, that item-level information alone would not supply the separate segment-completion observation. R2 needs the applicable IsPartial value. This follows from the different field meanings; it is not an additional invented event in AS64.

References: [ASR-ITEM](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Item.html), [ASR-RESULT](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Result.html)

## Choose the return that answers the required caption or export

Suppose one offer supplies only the completed E3 record. That return answers R2 for S1. It leaves unsubstantiated which explicitly nonrevisable items were available at E1 or E2. The eventual complete text cannot reconstruct those earlier flags.

A return supplying E2's selected Items and explicit flags answers R1 at E2. It does not yet answer R2. A return supplying the associated E1, E2 and E3 records answers both questions in this finite case. The additional value is identified early content, rather than an assumed speed advantage.

For an actual comparison, have the appointed application or platform owner identify the selected operation, applicable documentation and supported feature use, with evidence that stabilization is enabled. Connect the supplied items to the selected channel, alternative and segment using a supported association for that offered system. Then ask which actual flags answer each requested deliverable. An API name or transcript screenshot without the relevant properties leaves this comparison incomplete.

This requests an explanation of the offered records. It selects no request configuration, stability level, SDK, correlation algorithm or deployment procedure. AS64 contains no timing or capacity evidence with which to rank providers.

The related speech-input guide explains recording seconds, waveform samples and mel features. The CTC guide explains path and label probabilities. The streamed-response and watermark guides explain their own completion and revision boundaries. None of those quantities supplies the missing Stable evidence in this case.

## Connect a separately required equipment package

If the wider project independently requires approved equipment receiving, preparation or rack assembly, Leviathan's Rack Assembly & Preparation service and related project inquiry below provide a route to discuss that physical package.

Share a nonconfidential equipment summary, approved configuration, appointed supplier, proposed location, permitted physical tasks and required handoff to discuss scope, resources and availability. Recognition, application event handling, model integration, software hosting and commissioning, and transcript acceptance remain with the appointed software and application teams. The inquiry concerns the separately identified physical work.

## Primary sources

- [ASR-GUIDE — Amazon Transcribe — Streaming and partial results](https://docs.aws.amazon.com/transcribe/latest/dg/streaming-partial-results.html). Checked 2026-10-22. Latest developer guide: natural speech segments and item stability within partial output. Applies to the selected stabilization-enabled stream; AS64's event contents are fictional.
- [ASR-ITEM — Amazon Transcribe streaming API — Item](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Item.html). Checked 2026-10-22. Latest API reference: optional Stable Boolean, Content and pronunciation/punctuation Type values. Supports interpreting explicit item flags without supplying omitted values.
- [ASR-RESULT — Amazon Transcribe streaming API — Result](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_Result.html). Checked 2026-10-22. Latest API reference: Alternatives, ChannelId, IsPartial and ResultId. Segment associations must be established for the offered records; no cross-revision ID guarantee is inferred.
- [ASR-STREAM — Amazon Transcribe streaming API — StartStreamTranscription](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_StartStreamTranscription.html). Checked 2026-10-22. Latest operation reference identifies EnablePartialResultsStabilization. The case stipulates supported enabled use without recommending settings or adopting performance claims.

## Related Leviathan resources

- [Speech inputs: recording seconds, samples and mel features](/articles/speech-audio-samples-padded-mel-feature-workload)
- [CTC: best individual path and aggregated label probability](/articles/ai-ctc-frame-paths-collapsed-label-probability)
- [Streamed AI responses: content events and message completion](/articles/inference-stream-deltas-versus-completed-response)
- [Streaming aggregates: revisable values and finalized windows](/articles/ai-streaming-watermark-event-time-window-state)
- [Rack Assembly & Preparation for a separate equipment requirement](/services/gpu-assembly)
- [Discuss a separate receiving, preparation or rack-assembly project](/contact?service=gpu-assembly&from=%2Farticles%2Fstreaming-asr-item-stability-segment-finality)

## Does the wider project need physical equipment preparation?

Share a nonconfidential equipment and configuration summary, proposed location, permitted physical tasks and required handoff to discuss receiving, preparation or rack assembly.

[Discuss the scope](/contact?service=gpu-assembly&from=/articles/streaming-asr-item-stability-segment-finality)
