All blog posts
ModelsSTTTTSVoice AI

The State of Voice AI Models, Late 2026: What Changed in STT, TTS and Speech-to-Speech

October 16, 2026 · 7 min read · Ortavox Engineering

A builder's summary of the speech models released or updated in 2026 from OpenAI, ElevenLabs, Deepgram, Cartesia, Google, Mistral and the open-weights community, with verified dates and list prices, and what each change means for a voice agent in production.

2026 was the year the speech stack stopped being a bottleneck and became a menu. Streaming transcription with built-in turn detection is now a standard product from two vendors. Text-to-speech crossed into sub-100 ms model latency at more than forty languages. Speech-to-speech models got cheaper variants and a second generation. And the open-weights side produced models you can legitimately run in production under Apache 2.0. Here is what shipped, with dates and prices from vendor pages, and what it changes if you build voice agents.

Speech-to-text

Deepgram: Flux goes multilingual

Deepgram's Flux, the conversational model with end-of-turn detection built into the recogniser, became multilingual on April 29, 2026: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch, with mid-call language switching, through the cloud API or self-hosted. Nova-3 remains the general-purpose family for everything else, including diarisation and languages outside Flux's ten. Deepgram's own guidance is to start new voice agents on Flux and new transcription projects on Nova-3.

Cartesia: Ink-2 and turn detection in the STT

Cartesia released Ink-2 in May 2026 as a streaming STT model with built-in turn detection, English only at launch, then adjustable turn-start, turn-end and eager-end thresholds in June, keyterm prompting in July, and Spanish, French, Hindi and Japanese in September. Like Flux, it moves the end-of-utterance decision into the recogniser, which removes one model from the pipeline for latency-sensitive agents.

ElevenLabs: Scribe v2 family

Scribe v2 Realtime (streaming, around 150 ms partials per the product page) arrived in November 2025, with the batch Scribe v2 and a Scribe v2 Medical variant following in 2026. API list prices: $0.22 per audio hour for batch Scribe v2 and Scribe v2 Medical, $0.39 per hour for Scribe v2 Realtime.

OpenAI: streaming transcription as its own product

On May 7, 2026 OpenAI added gpt-realtime-whisper, a streaming speech-to-text model, and gpt-realtime-translate, a live translation model, to the Realtime API alongside GPT-Realtime-2. List prices are $0.017 per minute for live transcription (also offered as gpt-live-transcribe) and $0.034 per minute for live translation. Batch gpt-transcribe is $0.0045 per minute.

Google: a transcription-only Live model

Google's Live API now lists gemini-3.5-transcribe-live-preview, a speech-to-text-only model with interim and final transcripts, automatic language detection with code-switching, utterance timestamps and custom vocabulary biasing, in preview as of this writing. It sits beside the conversational Live models rather than replacing Cloud Speech-to-Text.

Mistral: two Voxtral transcription generations

Mistral shipped Voxtral Transcribe 2 in February 2026, with a batch model (Voxtral Mini Transcribe V2, $0.003 per minute) and Voxtral Realtime, released as open weights under Apache 2.0 and served at $0.006 per minute. On October 20, 2026 Voxtral 3 followed: Voxtral Small Transcribe 3 ($0.004 per minute) and Mini Transcribe 3 ($0.0035 per minute) for batch transcription and diarisation across 50+ languages, and Voxtral Mini Transcribe Realtime 3 at $0.008 per minute for agents. Mistral's Voxtral 3 page describes API access and self-deployment but does not state whether the generation-3 weights are public.

Streaming STT (list price)

Per minute

Turn detection built in

Deepgram Flux English / Multilingual

$0.0077 / $0.0078

Yes

Deepgram Nova-3 streaming mono / multi

$0.0077 / $0.0092

No

Cartesia Ink-2

3 credits per second (about $0.39 per hour on the Scale plan)

Yes

ElevenLabs Scribe v2 Realtime

$0.39 per hour (about $0.0065)

No

OpenAI gpt-realtime-whisper / gpt-live-transcribe

$0.017

No (VAD-driven)

Mistral Voxtral Mini Transcribe Realtime 3

$0.008

No

Mistral Voxtral Realtime (v2, Apache 2.0)

$0.006 hosted, or self-host

No

Text-to-speech

ElevenLabs Eleven v4

Eleven v4 and Eleven v4 Turbo launched on September 28, 2026 after a June preview, and were made available in Agents and the API on the same day. API list prices put v4 at $0.08 per 1K characters and v4 Turbo at $0.04, the same as v3 and v3 Conversational respectively, so the upgrade did not raise the price per character.

Cartesia Sonic 3.5 and 3.6

Cartesia moved fast: Sonic-3.5 previewed in April 2026 and went GA in May with speed and volume controls; Sonic-3.6 went GA on August 27, 2026 with Odia and Urdu added, for 44 languages and 61 locales, and a self-hosted release followed on September 17. The pricing model is credit-based: one credit per character, on plans from a free 20K credits to $299 per month for 8M.

Mistral Voxtral TTS

Mistral's first TTS shipped in March 2026 at $0.016 per 1K characters with open weights under CC BY-NC 4.0 and nine languages. Voxtral Mini TTS 3, in October, expanded to 35+ languages and 50+ dialects with emotion tags and pace settings, at $13 per million characters.

OpenAI and Deepgram

OpenAI's TTS line-up is unchanged on the price sheet: gpt-4o-mini-tts at $12 per million audio output tokens, tts-1 at $15 and tts-1-hd at $30 per million characters. Deepgram's Aura-2 lists at $0.030 per 1K characters.

Speech-to-speech

OpenAI released GPT-Realtime-2 on May 7, 2026 and gpt-realtime-2.1 with a 2.1-mini on July 6, 2026, reporting a p95 latency reduction of at least 25% across Realtime voice models and improvements to alphanumeric recognition, noise handling and interruption behaviour. Audio pricing for 2.1 is $32 per million input tokens, $0.40 cached and $64 per million output tokens; 2.1-mini is $10 and $20. A separate gpt-live-1 session product is billed at $0.05 per minute plus the backend model.

Google's gemini-3.8-live is generally available and recommended for low-latency multilingual voice agents, with native audio, proactive audio, affective dialog, tool use and a live-avatar capability; gemini-live-2.5-flash-native-audio remains GA alongside it. ElevenLabs lists a Realtime v1 Mini and Max for Agents at $0.08 per minute plus LLM usage. Mistral's Voxtral 3 launch bundled a voice-agent API built on LiveKit at $0.04 per minute with Mistral Small 4 for reasoning, or $0.08 with a larger model.

Open weights

Model

Type

Licence

When

Qwen3-TTS (0.6B and 1.7B; VoiceDesign, CustomVoice, Base)

TTS with 3-second voice cloning

Apache 2.0

January 22, 2026

Kyutai Pocket TTS (100M params, CPU real-time)

TTS

Open weights (see model card)

January 2026, more languages in April

Kyutai TTS 1.6B (used in Unmute)

Streaming TTS

Open weights (see model card)

2025, maintained

Mistral Voxtral Realtime (Transcribe 2)

Streaming STT

Apache 2.0

February 2026

Mistral Voxtral TTS

TTS

CC BY-NC 4.0

March 2026

NVIDIA Parakeet-TDT-0.6B-v3 and Canary-1B-v2

ASR / speech translation, 25 languages

CC-BY-4.0

2025, still current

The licence column is the one to read. Apache 2.0 and CC-BY-4.0 are usable commercially; CC BY-NC is not without a separate agreement. Ortavox already serves a self-hosted Kyutai TTS as a provider option, and the engine's provider registry is the mechanism for adding other open models when a customer's deployment calls for them.

What changed for builders

  1. Turn detection moved into the STT. With Flux and Ink-2, the recogniser emits end-of-turn events, which removes a model and its latency from the cascade. The Ortavox engine still runs its own turn-detection stage by default because it is provider-neutral, but delegated detection is the direction of travel.
  2. Streaming STT prices converged around $0.006 to $0.009 per minute at list, except OpenAI's $0.017. For a 10,000-minute month that spread is $60 to $170; it is rarely the number that decides a stack.
  3. TTS per-character prices held while quality and language counts rose. ElevenLabs kept v4 at v3 prices; Cartesia reached 44 languages; Mistral reached 35 languages for under two cents per thousand characters on its hosted model.
  4. Speech-to-speech got a cheap tier. gpt-realtime-2.1-mini at $10/$20 per million audio tokens makes a single-model agent viable for simple tasks, though you give up the ability to choose STT and TTS independently.
  5. Open weights became production-grade. Qwen3-TTS and Voxtral Realtime are Apache 2.0, which is what on-premise and data-residency deployments were waiting for.
⚠

List prices change often and most vendors run promotions. Every figure above is from the vendor's public pricing or documentation page as retrieved in October 2026; re-check before you budget.

Sources

  • OpenAI API pricing — https://developers.openai.com/api/docs/pricing
  • OpenAI Developer Community: New Realtime models gpt-realtime-2.1 — https://community.openai.com/t/new-realtime-models-on-the-api-gpt-realtime-2-1-and-gpt-realtime-2-1-mini/1385896
  • OpenAI Developers announcement of GPT-Realtime-2, Translate and Whisper (May 7, 2026) — https://x.com/OpenAIDevs/status/2052440968763515223
  • ElevenLabs API pricing — https://elevenlabs.io/pricing/api
  • ElevenLabs Scribe v2 Realtime — https://elevenlabs.io/realtime-speech-to-text
  • Unite.AI: ElevenLabs launches Eleven v4 — https://www.unite.ai/elevenlabs-launches-eleven-v4-with-low-latency-turbo-variant/
  • Deepgram pricing — https://deepgram.com/pricing
  • Business Wire: Deepgram launches Flux Multilingual (April 29, 2026) — https://www.businesswire.com/news/home/20260429043418/en/
  • Cartesia pricing — https://cartesia.ai/pricing
  • Cartesia changelog 2026 — https://docs.cartesia.ai/changelog/2026
  • Google Gemini Live API models — https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/live-api
  • Mistral: Voxtral 3 — https://mistral.ai/news/voxtral-3/
  • Mistral: Voxtral TTS — https://mistral.ai/news/voxtral-tts/
  • MarkTechPost: Voxtral Transcribe 2 (February 2026) — https://www.marktechpost.com/2026/02/04/mistral-ai-launches-voxtral-transcribe-2-pairing-batch-diarization-and-open-realtime-asr-for-multilingual-production-workloads-at-scale/
  • Qwen3-TTS repository — https://github.com/QwenLM/Qwen3-TTS
  • Kyutai TTS — https://kyutai.org/tts/
  • Canary-1B-v2 and Parakeet-TDT-0.6B-v3 (arXiv) — https://arxiv.org/abs/2509.14128

Ready to build?

Start with 100 free minutes. No credit card required.

Get started free