DictaFlow
DictaFlow Blog

Soniox STT v6 Migration: Test Dictation Output in 2026

API compatibility, currency wording and the application tests that belong between a new model and your users.

Download free for WindowsFree plan to start. One account on Windows, Mac, iPhone and Android.

November 04, 2026

Original microphone illustration branching into two differently formatted abstract output cards
Original procedural editorial illustration of audio routed to different text outputs. Not a product interface or benchmark result.

Soniox STT v6 migration needs a dictation-output check, even when the API request still works. Test currency wording, code hyphens and spoken editing commands before changing your default model. Soniox's November 3 changelog documents all three output behaviors, alongside API compatibility.[1]

The small formatting section caught my attention before the voice-agent features did. A dictation app can receive a perfectly valid response and still put the wrong kind of text into a document. That is the upgrade question here: does the finished text still match what your app promises?

There is fresh activity around that question today. A November 4 VivaDicta issue requests a v6 upgrade and calls out formatting tests.[2] MarkTechPost covered the release on November 3; Yowox followed on November 4.[3][8] The coverage is right to examine conversational context and echo. For a dictation app, I would start elsewhere: the output contract. Neither news attention nor a compatibility statement establishes that a particular formatter is ready. These reports are not evidence of widespread failures. I haven't benchmarked either model or tested VivaDicta.

Read compatibility as a starting point

Soniox lists stt-rt-v6 for streaming and stt-async-v6 for files. Its migration instruction is to change the model name.[1] That answers how to request the model. It doesn't settle how your application should handle the response.

Imagine a small repair shop building a voice-entry tool. Someone dictates a replacement-part note, a quoted amount and a request for the next shift. The shop needs searchable part identifiers. Its customer email needs normal sentences. The same spoken material can require different output in those two places.

An upgrade review should name those expectations before anyone compares transcripts. Otherwise, a reviewer may prefer whatever looks tidier, while another expects the speech preserved. Neither preference tells the developer which behavior belongs in the released app.

Write the expected text beside each test recording. Keep a second column for acceptable variations. A comma in a casual note might be flexible. A separator in an identifier might not be. These are proposed tests, not measured v6 results.

Make the amount test deliberately boring

The official example keeps "15 dollars and 20 cents" as spoken, instead of turning it into "$15.20".[1] That's useful evidence about currency wording. It isn't a promise about every number in every language.

For the hypothetical shop, record one short sentence about a quote. Give it two destinations: an email draft and a dedicated amount field. Decide separately what each destination accepts. Don't assume a transcript that reads well will also fit the field.

For an email, the spoken currency words may be perfectly usable. For a numeric field, your application may need an explicit conversion step. If that step already exists, inspect what it receives after the model change. If it doesn't exist, don't disguise the gap with a broad instruction to make everything look professional.

Keep the original transcript visible during this test. Compare it with the converted value and the saved record. If the amount changes, you need to know where it happened. A screenshot of the final field alone won't tell you whether recognition, conversion or entry introduced the difference.

I would start with invented amounts, not customer records. Add a case where the speaker restarts the amount halfway through. Decide whether the application should preserve the restart or interpret it as a correction. That decision belongs in the test expectation before the result arrives.

Commands are also ordinary words

Soniox says spoken commands such as "period", "new line" and "scratch that" are returned as words for the application to apply.[1] The recognizer passing them through doesn't itself establish that the app will edit the document.

That distinction deserves a paired test. First, dictate an instruction intended to create a line break. Then dictate a sentence discussing the phrase "new line". Label the intended result in advance. A formatter that consumes the phrase everywhere may damage the second sentence. One that never consumes it may disappoint users in the first.

Use the same approach for corrections. In a disposable note, deliberately abandon a phrase, then correct it. In another note, discuss the correction command as ordinary content. Check both. These examples test the app's interpretation rules; they don't establish how Soniox behaves beyond its documented transcript contract.

Avoid stacking a new recognizer and a rewritten cleanup prompt into the same experiment. Compare the recognizer change first. Then test a formatter change separately if needed. Otherwise, an improvement in one stage can conceal a regression in the other, and you won't know which change to keep.

For identifiers, include one code whose hyphen is spoken and another where it isn't. The changelog says v6 only writes those hyphens when spoken.[1] Compare the final string character by character with the identifier you intended. Search for it in a test record afterward, rather than stopping when the note looks plausible.

A successful request is only one checkpoint

For a Soniox STT v6 migration, I would retain four artifacts from each approved test: the audio, the returned transcript, the application's formatted text and the saved destination value. Keep them together with the model name and app version.

That sounds fussy for a short sentence. It saves a vague support exchange later. "The amount is wrong" gives you very little to investigate. A reproducible case showing the transcript and the conversion output gives you a place to start.

Check both fresh settings and an upgraded existing profile. The VivaDicta issue proposes updating its saved-model migration map as well as its defaults.[2] That's a contributor's request, not proof that the change has shipped. It does raise a useful question for other app builders: what happens to people who saved an older choice?

Soniox says v5 requests will automatically route to v6 after December 15, 2026.[1] Leaving the old model name in place therefore isn't a permanent way to preserve old output. Plan the review before that transition, and document a fallback that will actually remain available.

If you only want to dictate, skip the API work

Most writers shouldn't need to maintain this test machinery. They need a short sample that resembles their work, and a clear way to report an output change.

DictaFlow, this article's publisher, offers desktop hold-to-talk dictation that inserts text at the active cursor.[7] That is a finished input workflow, rather than a raw transcription endpoint. It doesn't establish a Soniox integration or mean every formatting expectation will match by default.

If you're evaluating DictaFlow, try an ordinary note plus an invented amount or identifier that matters to your task. Read the inserted text before using it. If you're maintaining a Soniox app instead, keep the model comparison separate from the application formatter.

The useful release evidence is a saved test case with an agreed expected result. Start with the amount and identifier your users would notice immediately if either changed.

Disclosure: Published by DictaFlow, with AI-assisted research and drafting. We have not benchmarked these models. The repair-shop cases are proposed tests.

Sources

[1] Soniox models and November 3 changelog

[2] VivaDicta upgrade request, November 4

[3] Yowox release coverage, November 4

[7] DictaFlow desktop workflow

[8] MarkTechPost release listing, November 3 (full article unavailable during research)

Download free for WindowsFree plan to start. One account on Windows, Mac, iPhone and Android.