A multilingual model can be easier to deploy than dozens of separate language recognizers, but optional hints and text normalization can change what reaches the editor. Deepgram’s October 2026 expansion is a timely case study.
The model and precise endpoint
Deepgram documents Nova-3 multilingual as model=nova-3 with language=multi on its /v1/listen API. The October 23, 2026 change exposes multilingual recognition across 55 languages, including mid-conversation switches. Users may supply repeated language_hint parameters or allow detection. This is one model path with a documented language selection mechanism, not a claim that all legacy Nova variants gain the same capabilities.
Hints bias; they do not guarantee
If an interview is known to alternate between Turkish and English, passing both languages as hints can help the recognizer resolve ambiguous speech. The absence of hints asks the system to detect languages automatically. A wrong or incomplete hint list may bias the output, so compare with an unhinted run on a sample before committing a production setting. Keep a record of code-switch transitions, not just overall word error rate.
Speaker identity is a separate operation
Nova-3 recognition and Deepgram diarization are related but different features. A recognizer can transcribe multilingual words without knowing who said them. The diarization document describes a versioned batch v2 and streaming v1, with compatibility constraints. If a reporter needs speaker-labelled quotations, test both models together and do not trust label identity without listening to the boundary.
Assess language fairness rather than a single score
Create a representative language set with native speakers, accented speakers, noise and borrowed words. Calculate word error rate on conventional word-segmented languages and character error rate where more appropriate. Keep error slices for proper names, numbers and code-switch boundaries. Vendor claims about comparative accuracy are vendor claims unless independently reproduced on the same released model and clean reference transcripts.
When it fits
Nova-3 multilingual is attractive for mixed-language recordings and live meeting capture. Offline Whisper may suit restricted connectivity or custom data governance, while AssemblyAI Universal-2 offers a broad documented language inventory. The acceptance decision should use actual field recordings and regional deployment constraints, not the size of a headline language count.
Worked production example
Imagine a journalist recording Turkish speakers who occasionally use English corporate terms and Russian names. Run a sample through Nova-3 multilingual both with relevant hints and without hints, then inspect how often terms are transcribed in the intended script. Accuracy should be reviewed by source-language readers, because normalizing a proper noun incorrectly may change the identity or meaning of a quoted person. Compare performance on the actual handheld recorder, not a clean product demonstration supplied by the provider.
Practical acceptance gate
For a bilingual interview, compute errors separately on native-language spans and switch boundaries; a single whole-file score can hide unacceptable performance on the smaller language. Record exact hint parameters and endpoint region. If speaker labels are necessary, add versioned diarization and correct it by hand. The output should carry a ‘machine-generated, editor reviewed’ status until all material quotations and numbers have been compared with the waveform.
Region, plan and model-version notes
Documentation checked on 2026-10-25. Exact applicability: vendor endpoint-specific. Version: nova-3 (language=multi). Plan: not independently verified for a specific plan. Monetary prices and quotas are not asserted unless separately verified; check the canonical AI 10 pricing ledger before purchase or publication of a tariff.
Sources & Further Reading
- Deep Change — official documentation (checked 2026-10-25)
- Deep Nova — official documentation (checked 2026-10-25)
- Deep Diar — official documentation (checked 2026-10-25)