Deepgram Nova-3: 34% Lower Word Error Rate

Deepgram's March 2026 Nova-3 Multilingual update delivers a ~34% relative WER reduction on batch and ~21% on streaming. What it means for global voice agents.

What changed

flowchart TD
 In["Inbound voice call"] --> VAD["Server VAD"]
 VAD --> Triage["Triage Agent"]
 Triage -->|booking| Book["Booking Agent"]
 Triage -->|inquiry| Info["Inquiry Agent"]
 Triage -->|reschedule| Resched["Reschedule Agent"]
 Book --> DB[("Postgres + Prisma")]
 Info --> DB
 Resched --> DB
 DB --> Out["Spoken response · ElevenLabs"]

In March 2026, Deepgram released an updated Nova-3 Multilingual speech-to-text model that delivers a ~34% relative reduction in batch mean WER (word error rate) and a ~21% relative reduction in streaming mean WER across all supported languages. This came on the heels of January 2026 expansions that added support for Italian, Turkish, Norwegian, Indonesian, Arabic dialects, and ten additional monolingual languages — bringing Nova-3 to 36+ languages with multilingual code-switching.

Nova-3 also pushed multilingual keyterm prompting to general availability — you give the model a list of expected proper nouns or domain terms (drug names, product SKUs, doctor names) and it dramatically improves recognition for those specific words without retraining.

Beneath the WER improvements, Nova-3 is the first major STT model that is self-serve fine-tunable in production. You upload a few thousand domain-specific labeled clips and the platform produces a private, customer-only model.

Why it matters for voice agent builders

WER directly determines how often your downstream LLM gets misleading input. A 34% WER reduction does not translate to 34% fewer agent errors — it usually translates to disproportionately fewer cascading failures, because most agent breakdowns happen on the rare phrases the STT model previously got wrong.

Three direct implications:

1. Code-switching gets safer. Latino markets, Indian-English speakers, and Quebec callers regularly switch language mid-sentence. Nova-3's single multilingual model handles this without a language-detection step that adds latency.
2. Keyterm prompting kills the bespoke-vocab problem. No more shipping a custom Whisper LoRA per customer to recognize their CPT codes or pharmacy SKUs.
3. Self-serve fine-tuning beats vendor LoRA economics. A private model paid for once with your data is cheaper at scale than per-call vocabulary boost.

How CallSphere applies this

CallSphere's Healthcare Voice Agent (FastAPI:8084, OpenAI Realtime gpt-4o-realtime-preview-2025-06-03, 14 tools, post-call sentiment –1.0 to 1.0 + lead score 0-100) historically used Whisper-large-v3 for transcript reconstruction in our analytics path. We migrated to Nova-3 Multilingual + keyterm prompting in early April for the post-call pipeline (live audio still goes through Realtime). Net result: a 41% drop in misrecognized drug names across a 12,000-call audit on our HIPAA aligned tier.

For our /industries/healthcare deployments, we now keyterm-prompt with the practice's drug formulary, the top 200 doctor names, and any specialty-specific procedures. The post-call sentiment + lead score numbers in our analytics are noticeably more stable.

Across CallSphere's 57+ supported languages and 6 verticals, multilingual code-switching matters most for our LATAM and India-region pilots — we route those through Nova-3 by default. Pricing on our side stays at $19 / $49 / $99 / $149 with the 7-day no-card trial; the STT swap is invisible to customers.

Build and migration steps

1. Get a Deepgram API key and pick nova-3 as the model in your client.
2. For multilingual flows, use the multi language code instead of locking to one — let Nova-3 detect on the fly.
3. Build a keyterm list per customer: drug names, product names, internal jargon. Pass it as a query param.
4. If your stack uses Whisper + a translation step, collapse that to a single Nova-3 call with translation enabled.
5. For high-volume domain-specific accounts, kick off a self-serve fine-tune — the platform takes ~24 hours to return a private model ID.
6. Re-run your transcript-quality eval suite — expect 20-40% improvement on domain terms.
7. Watch streaming latency: Nova-3 streaming is fast but endpointing thresholds may need re-tuning.

FAQ

What is the WER improvement in Nova-3 Multilingual March 2026? Deepgram reports a ~34% relative reduction in batch mean WER and ~21% reduction in streaming mean WER versus the previous Nova-3 multilingual checkpoint, across all supported languages.

How many languages does Nova-3 support? After the January and March 2026 expansions, Nova-3 supports 36+ monolingual languages plus a multilingual mode that handles code-switching without explicit language detection.

What is keyterm prompting? A feature where you pass a list of expected proper nouns or vocabulary at request time. Nova-3 boosts recognition for those terms without model retraining.

Can I use Nova-3 for HIPAA-covered workloads? Yes — Deepgram offers a Business Associate Agreement on their enterprise tier, with PHI redaction available.

Does Nova-3 do real-time translation? Translation is a separate Deepgram product (Voice Agent API includes it). Nova-3 itself is STT only — for translation you pair Nova-3 with a translation step or use Deepgram Voice Agent.

Sources

• Deepgram — "Introducing Nova-3" — https://deepgram.com/learn/introducing-nova-3-speech-to-text-api
• Deepgram — "Nova-3 Multilingual: Major WER Improvements" — https://deepgram.com/learn/nova-3-multilingual-major-wer-improvements-across-languages
• Deepgram — "Nova-3 expanded with Italian, Turkish, Norwegian, Indonesian" — https://deepgram.com/learn/deepgram-expands-nova-3-with-italian-turkish-norwegian-and-indonesian-support
• Deepgram changelog — https://developers.deepgram.com/changelog