Technology
November 12, 2026 1011 words

Voice AI Roundup: ElevenLabs Scribe v3, DeepL Aure, Fish Audio Live 3 and More

Hams AI Tech
Professional Content Team
Expert in AI Prompts & Professional Tools

Voice AI had one of its busiest stretches ever in October and November 2026. Transcription became more accurate across more languages, real-time translation started preserving the speaker's own voice and emotion, and voice agents learned when to listen and when to speak. For professionals who spend their days in meetings, calls and consultations, these advances translate directly into time saved. Here are the most important releases and what they mean for your work.

ElevenLabs Scribe v3 and ElevenAI

On November 10, ElevenLabs released two new speech-to-text models:

    • Scribe v3, for recorded audio, makes 27% fewer word errors than Scribe v2 across 102 languages and improves speaker diarization, the ability to tell who said what.
    • Scribe v3 Live, for low-latency use, makes 43% fewer errors than Scribe v2 Realtime and adds speaker labels and context-aware turn detection.

Both are available in the ElevenLabs API and are now the default in the ElevenLabs app. Scribe v3 Live is also in ElevenAgents.

At its New York summit the next day, ElevenLabs opened free early access to ElevenAI, a workplace app with dictation, consent-based meeting notes powered by Scribe v3, per-attendee action items and live in-call assistance. It also previewed Interaction Engine, which manages turn-taking and routes each part of a conversation to a cheaper or frontier language model, and announced Memory and Routines for ElevenAgents, Scribe v3 Medical, Nemotron 3.5 Super VL inside ElevenAgents, a Partner Program and the ElevenLabs Marketplace.

DeepL Aure: Translation in Your Own Voice

On November 11, DeepL introduced Aure, its most lifelike voice AI model, now powering DeepL Voice. Aure continuously adapts translated speech to the speaker's tone, emotion, emphasis and pace in real time, in their own voice. DeepL says the voice is never stored as a voice profile. DeepL Voice works in meetings on Microsoft Teams, Zoom and Google Meet, in in-person conversations and through the DeepL API for Voice, in English, German, Spanish, French, Italian, Japanese, Portuguese and Korean.

For international law firms, consultancies and healthcare providers, natural-sounding real-time translation could change how cross-border meetings work.

Fish Audio Live 3: Full-Duplex Conversation

Fish Audio released Live 3, a full-duplex voice model for voice agents. Full duplex means it keeps listening while it speaks. Trained on real conversations, it learns when to acknowledge with an mm-hmm, when to wait through pauses and corrections and when to stop if a caller interrupts. It can decide to respond in as little as 186 milliseconds. It is English only for now, with agent pricing of 0.035 dollars per minute through December 31, 2026.

Rime Prism: Inspectable Speech-to-Speech

Rime launched Prism in late October, a speech-to-speech model for voice agents that writes an inspectable text version of each reply before speaking. That design lets guardrails check every word while the audio keeps natural cues such as hesitation. Rime says it answers in about 300 milliseconds and runs tool calls in parallel with speech. Pricing on its pay-as-you-go plan is 0.080 dollars per minute, and the free plan includes 50 dollars of usage per month.

Open Speech Models

Open models made big gains too:

    • Salesforce SOV-Transcribe-1, a 26 billion parameter speech recognition model under Apache 2.0, which Salesforce says ranks first among open-weight models for non-streaming speech-to-text on Artificial Analysis.
    • NVIDIA Nemotron 3.5 Transcribe, a promptable model for transcription in 40 languages, speaker attribution, translation and audio questions, covered in our Nemotron article.
    • Mistral Voxtral 3, with transcription, emotional text-to-speech and voice cloning, covered in our Mistral article.
    • ServiceNow Apriel Voice 27B Preview, an open speech-language model that hears spoken requests directly and returns text and tool calls.
    • Zyphra ZONOS3-Mini, a compact voice-cloning model that runs in real time on a laptop.

What This Means for Professionals

Meetings and Consultations

Better diarization and lower error rates mean meeting notes that need less correction. Combine transcription with AI summaries to produce action items in minutes. Claude also now joins calls directly, covered in our Claude Meetings article. Always obtain consent before recording.

Legal Work

Depositions, client interviews and hearings can be transcribed and searched quickly, though certified transcripts still require qualified reporters where rules demand them. Lawyers can process transcripts with structured tools such as the AI cross-examination generator and trial preparation resources.

Healthcare

Medical-specific models such as Scribe v3 Medical aim at clinical documentation. Clinicians should use approved, compliant tools for patient data. Browse mental health resources for related workflows.

Customer Service

Full-duplex agents and inspectable speech-to-speech make voice bots sound more natural and safer to deploy. Start with simple tasks such as appointment scheduling.

International Teams

Real-time voice translation can reduce the need for interpreters in routine meetings, though important negotiations and legal proceedings still deserve professional interpreters.

Choosing a Voice Tool

    • For recorded meetings, compare Scribe v3, Nemotron Transcribe and SOV-Transcribe-1 on your own audio.
    • For live translation, test DeepL Voice with Aure in your common language pairs.
    • For voice agents, evaluate Fish Audio Live 3, Rime Prism and ElevenAgents on latency and interruption handling.
    • For privacy, consider open models you can host yourself.
    • For any option, test with recordings that reflect your real accents, terminology and background noise.

Frequently Asked Questions

How much better is Scribe v3?

ElevenLabs reports 27% fewer word errors than Scribe v2 across 102 languages.

Does DeepL store my voice?

DeepL says the voice is never stored as a voice profile.

What is a full-duplex voice model?

A model that listens while it speaks, allowing natural interruptions and acknowledgments.

The Bottom Line

Voice AI is becoming accurate, natural and affordable enough for everyday professional use. Start with meeting transcription, follow consent rules and expand to translation and voice agents as confidence grows. Measure accuracy on your own recordings before trusting any tool with important records, and keep a person responsible for reviewing anything that will be shared with clients or used as evidence. Explore our AI agent workflows and language AI resources for more.

Related AI News

Topics & Keywords
voice AI ElevenLabs Scribe v3 ElevenAI DeepL Aure DeepL Voice Fish Audio Live 3 Rime Prism speech to text