Best Voice AI Models in July 2026: STT, TTS, and Voice Agent Stack

By July 2026 the voice AI stack has plenty of production-grade options, so the hard part is choosing among them. Streaming STT runs under 300ms, real-time TTS answers in tens of milliseconds on vendor timing, and a fast LLM closes the loop. July’s news sat on the business side and the leaderboard: ElevenLabs opened talks for a tender offer near a $22 billion valuation, and the independent Artificial Analysis index kept ElevenLabs Scribe v2 the top generally available model at 2.2% WER. This guide picks the components of a production voice agent and the budget they must fit.

TL;DR: Best voice AI per layer, July 2026

If you only read one row: Deepgram Nova-3 plus Flux for STT, Cartesia Sonic-3.5 for TTS, a fast LLM such as Gemini 3.6 Flash or DeepSeek V4-Flash for the brain, and Retell or Vapi to orchestrate. That stack lands inside the sub-700ms practical round-trip and runs at production scale today.

The story of voice AI in July 2026

Voice AI spent the first half of 2026 turning research demos into datasheet-grade components, so July read as a month of business moves and leaderboard settling rather than a single headline model. Streaming STT sits under 300ms, real-time TTS answers in tens of milliseconds on vendor timing, and a fast LLM keeps the loop tight. The engineering question is which pieces compose for your accents, languages, and cost ceiling.

The loudest news was financial. On July 2, Bloomberg reported that ElevenLabs had opened early talks for a tender offer that would value the company near $22 billion, roughly double its February mark, with a reported $500 million in annual recurring revenue behind it. The talks are preliminary and structured as a secondary share sale, so treat the figure as a signal of momentum in the category rather than a closed round.

On measured accuracy, the independent Artificial Analysis word-error-rate index is now the cleanest cross-provider read. Among generally available models, ElevenLabs Scribe v2 leads at 2.2% WER, Microsoft MAI-Transcribe-1.5 follows at 2.4%, and AssemblyAI Universal-3.5 Pro lands at 3.1%. A non-GA research preview posts a lower error rate but is not deployable, so Scribe is the shipping accuracy leader. OpenAI’s GPT Transcribe joined the same board at 3.3%, and Deepgram Nova-3 measures 5.2% there while staying the streaming default on latency. These are batch numbers, so read them as accuracy potential, not streaming behavior.

Two shipping updates rounded out the month. AssemblyAI launched a Sync API on July 14 for synchronous short-form transcription, billed at $0.45 per hour on Universal-3.5 Pro, which suits request-response flows that cannot wait on a polling loop. Microsoft added MAI-Voice-2 Flash to Foundry as a faster variant of its expressive MAI-Voice-2 TTS, extending the June Build wave into a lower-latency option.

On TTS the shape of the market held. Cartesia Sonic-3.5 stays the latency pick, ElevenLabs splits into Flash v2.5 for real-time and Eleven v3 for expressive non-real-time work, Deepgram Aura-2 carries the enterprise HIPAA path, and Hume Octave 2 owns emotional control. Orchestration is the same four-way choice as prior months: Retell and Vapi as managed defaults, LiveKit Agents and Pipecat for the permissive-license self-host route, and OpenAI’s Realtime API for teams that want reasoning inside the audio loop.

The round-trip target still anchors on ITU-T G.114, which sets one-way mouth-to-ear delay at 150ms preferred and 400ms tolerable. Most production voice agents stay usable up to about 700ms round-trip. Hitting that budget in July 2026 is a matter of picking components that compose, then measuring your real p50 and p95 on your own traffic.

Best speech-to-text (STT) models in July 2026

Deepgram Nova-3. The streaming STT default

The right pick for any production voice agent that needs low-latency streaming transcription. Nova-3 is tuned for the sub-300ms budgets voice agents live in, and it pairs directly with Flux so end-of-turn detection runs in the same stack rather than as an afterthought.

Specs:

• WER: ~5.26% batch on Deepgram’s own benchmark; streaming ~6.84%; 5.2% on Artificial Analysis
• Latency: sub-300ms streaming (independent tests)
• Languages: streaming plus broad batch coverage
• Pricing: $0.0048/min streaming (pay as you go)
• Pairs with Deepgram Flux for end-of-turn detection

Best for: Production voice agents where end-to-end latency is the binding constraint. Live captioning. Real-time conversational AI.

Skip if: You want the lowest independently measured WER (use ElevenLabs Scribe v2). You need bundled diarization and keyterm prompting on the transcript (use AssemblyAI Universal-3.5 Pro Realtime).

ElevenLabs Scribe v2. The accuracy leader

The pick when transcription accuracy is the priority. Scribe v2 leads the generally available field on the independent Artificial Analysis word-error-rate index, and its Realtime sibling keeps that accuracy inside a live-agent latency window.

Specs:

• WER: 2.2% on Artificial Analysis (independent), top generally available model
• Latency: Scribe v2 Realtime answers in under 150ms (vendor)
• Languages: 90+ supported
• Pricing: ~$0.22/hr batch on Artificial Analysis; Scribe v2 Realtime at $0.39/hr standard ($0.28/hr on annual Business)

Best for: Voice agents in accent-heavy or high-stakes domains where a transcription error cascades into a wrong action. Meeting and medical scribing. Accessibility.

Skip if: You are standardized on Deepgram tooling and want Flux turn detection in the same stack (use Nova-3).

Microsoft MAI-Transcribe-1.5. The hyperscaler entry

The first-party cloud STT that lands near the top of the independent board. MAI-Transcribe-1.5 posts 2.4% WER on Artificial Analysis, third overall, and runs inside Microsoft Foundry with keyword biasing for domain terms.

Specs:

• WER: 2.4% on Artificial Analysis (independent), #3 overall
• Languages: 43
• Throughput: transcribes an hour of audio in well under a minute
• Keyword biasing for domain vocabulary
• Pricing: ~$0.36/hr ($6 per 1,000 minutes) in Microsoft Foundry

Best for: Teams already on Azure or Foundry, batch and near-real-time transcription at scale, multilingual workloads that benefit from keyword biasing.

Skip if: You need the tightest streaming turn-taking loop today (pair Nova-3 with Flux).

AssemblyAI Universal-3.5 Pro Realtime. Streaming with structure

The pick when the transcript needs structure, not just words. Universal-3.5 Pro Realtime (u3-rt-pro) brings live speaker diarization, keyterm prompting, disfluency control, and code-switching to streaming, and the batch model measures 3.1% WER on Artificial Analysis.

Specs:

• WER: 3.1% on Artificial Analysis (independent) for the batch Universal-3.5 Pro
• Real-time diarization and keyterm prompting on the streaming transcript
• New Sync API (July 14) for synchronous short-form transcription
• Pricing: $0.45/hr streaming (u3-rt-pro); $0.21/hr batch (universal-3-pro); Medical Mode +$0.15/hr

Best for: Multi-speaker calls, domain jargon that generic models mistranscribe, and products that need speaker labels live rather than after the call.

Skip if: Pure lowest-latency transcription is all you need (use Deepgram Nova-3).

OpenAI GPT Transcribe. The OpenAI-ecosystem pick

The default when your product already lives in the OpenAI API. GPT Transcribe (gpt-transcribe) joined the Artificial Analysis board at 3.3% WER, and the mini variant trades a little accuracy for a lower rate, both integrating cleanly with the rest of the OpenAI audio stack.

Specs:

• WER: 3.3% on Artificial Analysis (independent); mini variant at 4.5%
• Languages: broad multilingual coverage
• Latency: near real-time; the Realtime path answers under 150ms
• Pricing: ~$0.27/hr GPT Transcribe ($4.50 per 1,000 minutes); ~$0.18/hr for the mini variant

Best for: Teams building on OpenAI who want one vendor for transcription, LLM, and speech-to-speech.

Skip if: You need the lowest independent WER (Scribe v2) or model-integrated turn detection (Flux).

Deepgram Flux. The turn-taking layer

Not a standalone transcriber, but it solves a problem the rest of the category leaves to the application. Generic STT APIs return a transcript and stop. Flux models when the speaker has actually finished, and its eager end-of-turn signal can fire 150 to 250ms earlier than waiting for a full pause.

Specs:

• Model-integrated end-of-turn detection, runs with Nova-3
• Eager end-of-turn firing ahead of a full pause
• Pricing: $0.0065/min EN

Best for: Any turn-based production voice agent. The difference between an agent that interrupts naturally and one that talks over users or stalls.

Skip if: Your product is dictation or transcription only, with no turn-taking.

Best text-to-speech (TTS) models in July 2026

TTS is optimization on three independent axes: latency, naturalness, and emotional control. They do not move together, so pick by your dominant constraint. One caution up front: vendor time-to-first-audio numbers are best-case, and independent p50 streaming latency runs meaningfully higher.

Cartesia Sonic-3.5. The latency leader

The structural pick when round-trip latency is the binding constraint. Cartesia Sonic-3.5 posts vendor time-to-first-audio under 90ms, which leaves budget for the rest of the stack, and it streams output with voice cloning support.

Specs:

• TTFA: sub-90ms (vendor); higher on independent p50 streaming benchmarks
• Streaming output with voice cloning
• Broad multilingual coverage
• Pricing: credit-based tiers

Best for: Sub-500ms round-trip voice agents, telephony where latency drives perceived call quality, real-time interactive applications.

Skip if: Maximum expressive range matters more than latency (use Eleven v3 or Hume Octave 2).

ElevenLabs Flash v2.5. The conversational pick

The real-time ElevenLabs path. Flash v2.5 answers near 75ms model inference, which keeps ElevenLabs voice quality inside a live-agent budget without reaching for the slower Eleven v3.

Specs:

• TTFA: ~75ms model inference (vendor); higher on independent p50 streaming benchmarks
• Languages: 32
• Voice cloning supported
• Pricing: pay as you go, character-based

Best for: Conversational agents that want ElevenLabs voices with real-time latency, branded consumer voice products.

Skip if: You need the lowest possible time-to-first-audio (use Cartesia Sonic-3.5).

Eleven v3. Expressive quality, not real-time

The most expressive ElevenLabs model, positioned for quality rather than latency. Eleven v3 uses Audio Tags for emotional control and covers 70+ languages, and ElevenLabs itself points real-time and conversational use cases back to Flash v2.5.

Specs:

• Audio Tags for granular emotional control
• Languages: 70+
• High latency by design; not real-time-optimized
• Pricing: media / expressive tier

Best for: Audiobooks, branded voice, character work, and any output rendered ahead of time.

Skip if: The voice runs inside a live agent turn (use Flash v2.5).

Deepgram Aura-2. Enterprise TTS with HIPAA

The pick when you want Deepgram’s stack end to end with a compliance path. Aura-2 ships a broad voice set and a signed BAA on Enterprise, which matters for regulated deployments that already run Nova-3 and Flux.

Specs:

• Latency: sub-200ms vendor steady-state; higher on independent p50 streaming benchmarks
• Voices: 40+ across supported languages
• Compliance: HIPAA BAA on Enterprise
• Pricing: $0.030 per 1,000 characters (pay as you go)

Best for: Healthcare and regulated voice agents already using Deepgram Nova-3 and Flux, teams that want one vendor for STT and TTS.

Skip if: You need the broadest multilingual coverage (Cartesia or ElevenLabs cover more languages).

Hume Octave 2. The emotion specialist

The emotion-first TTS pick. Hume Octave 2 is built for prosody and emotional expression, with voice conversion and phoneme-level editing on top.

Specs:

• Latency: under 200ms
• Languages: 11
• Voice conversion and phoneme-level editing
• Pricing: dedicated capacity; under 1 cent per minute at volume

Best for: Mental-health products, character voices for games, emotion-sensitive content.

Skip if: Latency or broad language coverage is the primary constraint.

Microsoft MAI-Voice-2 and MAI-Voice-2 Flash. The hyperscaler expressive path

The Foundry TTS line. MAI-Voice-2 is Microsoft’s expressive model with granular emotion tags and zero-shot voice cloning across 15 languages, and July added MAI-Voice-2 Flash as the faster variant for latency-sensitive paths.

Specs:

• MAI-Voice-2: expressive TTS, 15 languages, emotion tags, zero-shot voice cloning
• MAI-Voice-2 Flash: July Foundry addition, faster variant for lower-latency use
• 5-second voice cloning with system-level consent enforcement
• Pricing: Microsoft Foundry / Azure Speech

Best for: Azure and Foundry teams that want first-party expressive TTS, assistants and contact-center voices in the Microsoft ecosystem.

Skip if: You need a published, independently measured time-to-first-audio to budget against (use Cartesia or ElevenLabs for latency-critical paths).

OpenAI gpt-realtime. Speech-to-speech

The native speech-to-speech option. gpt-realtime-2.1 keeps reasoning inside the audio loop, which skips the separate STT and TTS steps for a single low-latency model priced per audio token.

Specs:

• Speech-to-speech with reasoning in the audio loop
• Priced per audio token rather than per minute
• Single-model loop, no separate STT and TTS steps

Best for: Products where prosody and reasoning matter more than per-minute price, and where a single-model loop simplifies the stack.

Skip if: Per-minute cost predictability is the priority (use a classic STT-LLM-TTS pipeline).

Best voice agent platforms in July 2026

If you do not want to wire STT, LLM, TTS, and orchestration yourself, the platforms below ship in days what custom builds ship in quarters. The compliance line moved this month: on the major managed platforms a signed HIPAA BAA now sits behind Enterprise or a paid add-on, so confirm the tier before you build rather than after.

Retell AI. The most-teams default

The right default for most production voice-agent teams. Retell lands all-in around $0.07 to $0.31 per minute once LLM, TTS, and telephony are included, with a no-code builder and a developer SDK on top.

Specs:

• Pricing: all-in ~$0.07-0.31/min depending on model, TTS, and add-ons
• Builder: no-code visual builder plus SDK
• Compliance: HIPAA and custom BAA on the Enterprise plan (not pay-as-you-go)
• Latency: sub-700ms round-trip achievable with fast components

Best for: Most production voice agents where sub-700ms is acceptable and a managed platform reduces engineering load.

Skip if: You need bring-your-own-component control at scale (use Vapi) or a fully self-hosted stack (use LiveKit or Pipecat).

Vapi. The scale and BYO pick

The pick when you want to bring your own STT, LLM, and TTS and run at volume. Vapi charges a $0.05 per minute platform fee on top of component passthrough, landing all-in around $0.30 to $0.33 per minute.

Specs:

• Pricing: $0.05/min platform fee plus passthrough; all-in ~$0.30-0.33/min
• Bring-your-own STT / LLM / TTS
• Multi-channel: voice, plus SMS and chat
• Compliance: HIPAA add-on at $2,000/mo; Zero Data Retention at $1,000/mo

Best for: Teams that want component-level control, multi-channel deployments, and volume scale.

Skip if: You want the lowest total cost of ownership under 100K minutes/month (Retell’s base usually wins) or a self-hosted stack (LiveKit or Pipecat).

OpenAI Realtime API. Native speech-to-speech

The pick when you want one model for the whole audio loop. The Realtime API runs gpt-realtime speech-to-speech, priced per audio token rather than per minute.

Specs:

• Model: gpt-realtime, speech-to-speech
• Pricing: per audio token, with cached input discounted
• No separate STT and TTS steps

Best for: Prosody-sensitive products, single-vendor OpenAI stacks, and teams that value a simpler loop over per-minute predictability.

Skip if: You need per-minute cost control or self-hosting.

LiveKit Agents. Open-source orchestration at scale

The pick when you want to own the orchestration and still get first-party adapters for every STT and TTS. LiveKit Agents is Apache-2.0 and free to self-host, with a managed Cloud on top.

Specs:

• License: Apache-2.0 (self-host free)
• Cloud tiers up to Scale and Enterprise
• Compliance: SOC 2 Type II from the Scale tier; signed HIPAA BAA on Enterprise only
• First-party adapters for major STT and TTS providers

Best for: Teams that want open-source orchestration with production infrastructure and broad provider support.

Skip if: You want a fully managed no-code platform (use Retell) or a signed HIPAA BAA below Enterprise.

Pipecat. The open-source framework

The pick for a fully open, self-composed STT-LLM-TTS loop. Pipecat is BSD-2-Clause with a strong plugin ecosystem for Cartesia, Deepgram, ElevenLabs, and OpenAI, plus a managed Pipecat Cloud.

Specs:

• License: BSD-2-Clause (framework free)
• Managed option: Pipecat Cloud
• Plugin ecosystem across major STT / TTS / LLM providers
• Compliance: whatever your own deployment provides

Best for: Teams that want full control of the orchestration code, retry logic, and barge-in handling.

Skip if: You want a managed platform to own reliability and on-call for you (use Retell or Vapi).

HIPAA tier matrix

HIPAA support is the cleanest differentiator across the managed platforms. Signed BAAs sit behind Enterprise or a paid add-on on all of them in July 2026, so this single matrix often decides the pick for regulated teams:

Vendor latency vs independent benchmarks

Every TTS vendor publishes a best-case time-to-first-audio, and those numbers do not match what you measure in production. Independent p50 streaming latency runs meaningfully higher than the vendor figures, so the vendor number is a floor, not what your users hear. On the STT side, self-reported WER also runs optimistic against the neutral Artificial Analysis index, and batch accuracy does not predict streaming behavior. Run a domain reproduction with your accents, your background noise, and your prompts before you commit.

End-to-end latency budget. The math

Voice agents have a hard latency target. The ITU-T G.114 one-way mouth-to-ear recommendation is 150ms preferred and 400ms tolerable, and that constraint shapes any voice-agent budget. Sub-500ms round-trip is the aggressive target, and sub-700ms is the threshold most production use cases accept. Hitting either requires component picks that compose to the budget.

The breakdown:

The chart above uses vendor best-case time-to-first-audio to show the component budget. Independent p50 streaming latency runs meaningfully higher, so a real deployment sits nearer the top of each range. That gap between vendor timing and measured p50 is exactly why the last step of any voice build is to profile your own p50 and p95 rather than sum the datasheet numbers. For the LLM slot, a fast model keeps the loop tight; see the July LLM guide for the current low-latency picks.

Cost at scale: what 100K minutes/month actually costs

Per-minute list price hides the real production cost, and by July 2026 the paths split into managed flat-fee, bring-your-own with passthrough, native token-metered audio, and self-hosted with engineering load. The two figures that clear an independent check are the all-in aggregator estimates for Retell and Vapi:

The honest framing holds from prior months: under roughly 100K minutes/month, Retell’s base usually wins on total cost of ownership because the engineering time saved on passthrough tuning, retry logic, and compliance paperwork dominates the per-minute delta. Above about 1M minutes/month, bring-your-own or self-hosted economics start to flip, and the crossover depends on your retry rate and engineering cost. The OpenAI Realtime path is the premium option and the right pick only when prosody and a single-model loop matter more than per-minute price.

Decision framework

Choose Retell AI if:

• You are building most-team production voice agents.
• Sub-700ms round-trip is acceptable.
• You want a managed platform with a no-code builder and an SDK.
• You can move to the Enterprise plan when you need a signed HIPAA BAA.

Choose Vapi if:

• You want bring-your-own STT, LLM, and TTS control.
• You need multi-channel (voice plus SMS and chat).
• You are running at volume and can absorb the $2,000/mo HIPAA add-on if regulated.

Choose OpenAI Realtime API (gpt-realtime) if:

• You want native speech-to-speech with reasoning in the audio loop.
• Prosody matters more than per-minute price predictability.
• A single-vendor OpenAI stack simplifies your build.

Choose LiveKit Agents or Pipecat if:

• You want permissive open-source orchestration (Apache-2.0 or BSD-2-Clause) with no platform fee.
• You are willing to own reliability, retries, and on-call.
• You need first-party adapters across many STT and TTS providers.

Roll your own (Cartesia + Deepgram + a fast LLM) if:

• Sub-500ms round-trip is the target.
• You have the engineering team to build and run orchestration.
• The managed platforms do not support your specific stack.

Common mistakes when picking voice AI components in July 2026

1. Budgeting from vendor best-case latency. A sub-100ms vendor time-to-first-audio is a floor, and independent p50 streaming benchmarks put it meaningfully higher. Budget from measured p50, not the datasheet.

2. Reading batch WER as streaming accuracy. The Artificial Analysis numbers are batch. A model at 2.2% batch does not guarantee 2.2% on your streaming traffic, so validate streaming behavior on your own audio.

3. Skipping turn-taking detection. Generic STT APIs stop at the transcript. Without Flux or an equivalent end-of-turn layer, agents talk over users who pause or sit silent when they stop.

4. Treating the LLM as free latency. A 300ms STT paired with a 1,200ms LLM is not a fast agent. Pick a fast model for sub-500ms targets and measure the whole loop.

5. Reading HIPAA as included. In July 2026 Retell, Vapi, LiveKit, and Deepgram all gate a signed BAA behind Enterprise or a paid add-on. Confirm the compliance line before you build, not after.

How Future AGI fits

Voice agents fail in production for the same reasons text agents do: hallucinations, retry loops, accent edge cases, off-policy responses, and prompt injection through transcript contamination. Future AGI ships the eval, simulate, and observability layer that voice teams pair with their framework of choice:

• Simulate generates voice scenarios (accents, background noise, interruptions, ambiguous phrasing) and replays them against your agent before you ship.
• Evaluate scores every turn on groundedness, hallucination, tool-call accuracy, and accent handling, with voice-specific axes that text-only evals miss.
• Agent Command Center applies runtime guardrails that block bad outputs at the gateway in low hundreds of milliseconds, which fits inside the voice latency budget.
• Error Feeds cluster live failures so you see “accent-X failing on intent-Y” instead of dozens of unrelated tickets.
• Optimize auto-rewrites prompts and policies, then re-validates against your regression set.

Future AGI is a companion to Vapi, Retell, LiveKit, and Pipecat rather than a competitor on the voice-framework axis. For voice specifically, the eval suite adds accent-handling, sentiment-consistency, and tool-call-accuracy checks that a text-only harness never surfaces.

July 2026 settled into two truths: the money is chasing the category, and the leaderboard now separates the picks cleanly by accuracy, latency, and compliance rather than by whether a layer works at all. Every layer has at least two production-grade options, so the stack you assemble matters less than the budget it lands inside and the measurement loop you wrap around it.

Treat a voice agent as a latency, cost, and reliability system rather than a favorite-model bet. Shortlist by your binding constraint, wire the STT-LLM-TTS-orchestration path, then measure real p50 and p95 on your own accents and noise before you ship. That eval loop, not the leaderboard row, is what turns July’s mature components into an agent that holds up on a live call.

Sources

STT primary

• Deepgram Nova-3 announcement (~5.26% batch WER)
• Deepgram Flux conversational STT (model-integrated end-of-turn)
• Deepgram pricing
• AssemblyAI pricing (Universal-3.5 Pro, u3-rt-pro, Sync API)
• AssemblyAI speech-to-text products
• ElevenLabs Scribe v2 Realtime (under 150ms)
• Microsoft MAI-Transcribe-1.5
• OpenAI voice models in the API

TTS primary

• Cartesia Sonic-3.5 launch
• Cartesia TTS models docs
• ElevenLabs models
• ElevenLabs Flash v2.5 latency (~75ms model inference)
• Eleven v3 (Audio Tags, 70+ languages, expressive)
• Deepgram Aura-2 enterprise TTS
• Hume Octave 2 (under 200ms, 11 languages)
• Microsoft MAI-Voice-2
• Microsoft MAI-Image-2.5 Pro and MAI-Voice-2 Flash in Foundry

Voice agent platforms

• Retell AI pricing
• Vapi pricing
• LiveKit pricing and tiers
• LiveKit Agents docs
• Pipecat

Business and independent benchmarks

• Bloomberg: ElevenLabs in talks for tender offer at $22B valuation
• Artificial Analysis Speech-to-Text (independent WER)
• Artificial Analysis Text-to-Speech Arena (quality Elo)
• ITU-T Recommendation G.114 (one-way transmission time)

See also: Best LLMs of July 2026 for the LLM brain in your voice agent. Previous voice post: Best Voice AI of June 2026.