Is Gradium a good Cartesia alternative for voice agents?

For real-time voice agents, both platforms clear the practical TTFA threshold of 300 ms with streaming delivery. Cartesia reports 90 ms TTFA on Sonic-3. Gradium publishes a full benchmark with P50 TTFA of 214 ms excluding connection establishment, alongside ElevenLabs, Mistral Voxtral, and OpenAI GPT-4o Mini comparisons. Where Gradium stands out is voice-agent-tuned pronunciation, natural turn-taking through semantic VAD, accent-preserving voice cloning, and an integrated TTS and STT stack from a single provider.

How does Gradium compare to Cartesia on TTS latency?

Both platforms report TTFA figures well under the 300 ms streaming threshold that makes voice agents feel natural. Cartesia reports 90 ms time-to-first-audio on Sonic-3. Gradium publishes a full benchmark: P50 258 ms, P95 274 ms end-to-end, and P50 214 ms, P95 228 ms when excluding connection establishment (measured from Paris, 15-25 word sentence, WebSocket, 100 queries, warm state). Gradium's benchmark also includes matched-methodology comparisons against ElevenLabs Turbo v2.5, ElevenLabs Flash v2.5, Mistral Voxtral TTS, and OpenAI GPT-4o Mini. For voice agents, both vendors are a strong fit on latency; the differentiators to weigh are voice-agent-tuned pronunciation and semantic VAD.

Why is Gradium TTS better suited to voice agents than general-purpose TTS?

Voice agents have to speak inputs that general-purpose TTS models handle poorly: phone numbers, URLs, email addresses, dates, times, street addresses, order IDs, confirmation codes, dollar amounts, and named entities. Gradium's TTS is tuned specifically for these cases. Phone numbers are grouped and intoned naturally rather than read as a flat digit string. URLs and email addresses are spelled out with correct handling of domains, dots, dashes, and slashes. Dates and times follow the regional convention of the target language. Named entities are pronounced consistently across a session.

What is semantic VAD and why does it matter when comparing Gradium to Cartesia?

Semantic voice activity detection determines not just when a speaker has gone silent, but when they have finished a complete thought. In conversational AI, this prevents the system from cutting in mid-sentence. Gradium's STT includes semantic VAD natively. Cartesia's Ink is positioned primarily as a transcription model, with turn-taking left to the surrounding pipeline.

Does Gradium support voice cloning like Cartesia?

Yes. Both Gradium and Cartesia offer instant voice cloning from 10 seconds of audio. Gradium also offers Pro Voice Clones, a fine-tuned model for highest fidelity, available from the M plan. Explicit consent is required before cloning any voice.

Can Gradium clone accents and speaking styles within a language?

Yes. Gradium's voice cloning reproduces the accent and the speaking style present in the 10-second reference sample, within each of its five supported languages, rather than defaulting to a single standard pronunciation. This covers the major regional accents in English, French, Spanish, German, and Portuguese, and speaking styles including conversational, narrative, broadcast, customer-service, expressive, and whispered delivery.

How many languages does Gradium support?

Gradium supports five languages with native fluency: English, French, Spanish, German, and Portuguese. It also supports mid-sentence code-switching between these languages with no latency penalty. Cartesia supports 40+ languages and is stronger for wide multilingual coverage.

What SDKs does Gradium offer?

Gradium provides official SDKs in Python and Rust, and integrates natively with LiveKit and Pipecat.

Can Gradium be deployed on-premise?

Yes. Gradium supports the full range of deployment surfaces: cloud marketplace, private cloud (tenant-isolated in the customer's account), on-premise (including air-gapped), and on-device (for latency-critical or offline use cases). The same model and API surface is available across all four, so a deployment can move from cloud to on-prem or on-device without re-architecting. Enterprise plans also include zero data retention and SLA commitments. Cartesia is certified SOC 2 Type II, HIPAA, and PCI Level 1.

Does Gradium offer a free plan?

Yes. Gradium's Free plan includes 45,000 credits per month (approximately 1 hour of TTS or 4 hours of STT), 5 Instant Voice Clones, and full API and Studio access. Commercial use is not included on the Free plan. Cartesia's Free plan includes 20,000 credits per month.

How does Gradium pricing work?

Gradium uses a credit system: 1 character of TTS equals 1 credit, and 1 second of STT equals 3 credits. Plans range from Free ($0/month, 45,000 credits) to Tailored (custom pricing). Additional credits can be purchased pay-as-you-go on any paid plan.

Gradium was founded by the co-founders of Kyutai, a research lab known for peer-reviewed work on audio language models and real-time voice AI.

Where can I get started with Gradium?

Gradium is available at gradium.ai. The Free plan ($0/month, 45,000 credits) is the fastest way to test the API. Integration guides are available for streaming TTS, streaming STT, and instant voice cloning.

Comparison

Cartesia Alternative: Gradium for Real-Time Voice AI

April 20, 202612 min read

Updated September 10, 2026

By Gradium. Data as of September 2026.

Key takeaways

This is the migration view for teams on Cartesia. For the wider field, see best ElevenLabs alternatives 2026.
Cartesia's strength is voice preference and language coverage. Sonic 3.6, released August 27, 2026, ranked first of 92 at 1,282 Elo on the Artificial Analysis provider-voice board read September 8, 2026.
On the Coval latency read of the same date, Sonic 3.6 measured 440 ms median perceived time to first audio with a 171 ms P25 to P75 spread; Sonic 3.5 measured 269 ms with 46 ms, and Gradium TTS 214 ms with 31 ms.
On the August 2026 hard-case pronunciation test, Gradium TTS passed 81.0% and Cartesia Sonic 3.6 75.1%.
Cartesia prices in credits ($5 for 100k on Pro, $299 for 8M on Scale) and added Speech-to-Text with Ink-2 on July 9, 2026.

Gradium is a real-time voice AI platform built by the co-founders of Kyutai. It offers streaming Text-to-Speech and streaming Speech-to-Text with semantic voice activity detection (VAD) over WebSocket, plus a REST voice cloning API that produces an instant clone from 10 seconds of audio. It is the closest direct alternative to Cartesia for teams building conversational voice agents, and the only one of the two that ships voice-agent-tuned TTS and semantic VAD natively in its STT.

Who is this for. Developers and technical teams evaluating Cartesia (Sonic 3.6 Text-to-Speech, Ink-2 Speech-to-Text, Line voice agent platform) and looking for a provider with robust pronunciation on voice-agent-specific inputs (phone numbers, URLs, email addresses, complex entities), integrated semantic turn-taking, and a unified streaming stack.

How Do Gradium and Cartesia Compare at a Glance?

Sources: vendor documentation and pricing pages, read September 8, 2026; Coval leaderboard, 1-day window, read September 8, 2026 (latency); Gradium hard-case evaluation, August 2026 (pronunciation pass rate).

Dimension	Gradium	Cartesia
Primary use case	Real-time voice agents and developer voice APIs	Real-time streaming applications, expanding into voice agents with Line platform
TTS model	TTS (streaming), designed for voice agents with robust handling of complex inputs (phone numbers, URLs, email addresses, dates, times, street addresses, order IDs, named entities)	Sonic 3.6, designed for real-time streaming applications
TTS latency (TTFA)	P50 258 ms, P95 274 ms end-to-end; P50 214 ms, P95 228 ms excluding connection establishment (published benchmark). Under the 300 ms streaming threshold for natural turn-taking	90 ms (vendor claim). Well under the 300 ms streaming threshold for natural turn-taking
Word-level timestamps (TTS)	Yes, high-precision	Yes
STT model	STT (streaming)	Ink-Whisper, 66 ms TTCT, included in all plans
Semantic VAD	Yes, included in the STT	Not documented as a core feature
Voice cloning	Instant (10 s of audio) + Pro (fine-tuned). Preserves a wide range of accents and speaking styles within each supported language	Instant (10 s of audio) + Pro
Voice library	Curated library of expressive voices	Curated voice library
Languages	English, French, Spanish, German, Portuguese, with regular updates	40+ languages
Mid-sentence code-switching	Yes, no latency penalty	Not documented
Agent framework integrations	LiveKit, Pipecat	Vapi, LiveKit, Pipecat
SDKs	Python, Rust (official)	Multiple official SDKs (see Cartesia docs)
Deployment options	Cloud marketplace, private cloud, on-premise, on-device, from the same model and API	Cloud SaaS; enterprise deployment options available; certified SOC 2 Type II, HIPAA, PCI Level 1
Enterprise data control	Zero data retention, SLA commitments (enterprise plans)	SOC 2 Type II, HIPAA, PCI Level 1
Free plan	$0/month, 45,000 credits	$0/month, 20,000 credits
Founders	Neil Zeghidour, Laurent Mazaré, Olivier Teboul, and Alexandre Défossez, co-founders of Kyutai	Karan Goel and Albert Gu (Stanford AI lab, Mamba architecture)

Who Is Cartesia?

Cartesia is a voice AI company whose product line includes Sonic (Text-to-Speech), Ink (Speech-to-Text) and Line (a voice agent platform). Sonic 3.6, released August 27, 2026, is its most recent Text-to-Speech model, designed for real-time streaming applications, with a reported sub-90 ms time-to-first-audio and support for 40+ languages.

Who Is Gradium?

Gradium is a real-time voice AI platform for developers and companies deploying voice agents. Its product surface is:

A streaming Text-to-Speech API over WebSocket
A streaming Speech-to-Text API with semantic voice activity detection
A voice cloning API, available as Instant (zero-shot, 10 seconds of audio) and Pro (fine-tuned model)

All three APIs share a streaming-first architecture on WebSocket, suitable for bidirectional, low-latency communication in production.

Gradium was founded by Neil Zeghidour, Laurent Mazaré, Olivier Teboul, and Alexandre Défossez, the co-founders of Kyutai, a research lab with peer-reviewed work on streaming speech models. Kyutai released world-first open systems including Moshi (real-time Speech-To-Speech) and Hibiki (live Speech-To-Speech translation). Gradium translates that research into production infrastructure.

What Should You Look for in a Cartesia Alternative?

Teams evaluating alternatives typically focus on the following criteria, grouped by TTS, STT, voice cloning, and platform concerns. Gradium was designed with all of them in mind.

TTS

Pronunciation robustness on complex inputs. Most TTS models fall short on the inputs that matter most in voice agents: phone numbers, URLs, email addresses, dates, times, street addresses, order IDs, named entities. Gradium's TTS is specifically tailored for these cases. You can further fine-tune pronunciation using custom pronunciations and text normalization rules.
TTFA latency at or below 300 ms, delivered over streaming. For real-time voice agents, a TTFA under 300 ms combined with a true streaming API is the practical threshold for natural-feeling turn-taking. Both Gradium and Cartesia meet that bar.
Streaming API. TTS should stream over WebSocket so audio can be delivered incrementally as it is generated. Both vendors support this. For high-throughput pipelines, Gradium also supports multiplexing multiple TTS requests over a single WebSocket connection.
Voice naturalness and quality, with a ready-to-use voice library. Production teams usually want expressive, natural-sounding voices available out of the box, without having to clone or tune for every project. Both Gradium and Cartesia ship curated voice libraries covering a range of speaker identities and styles.

STT

Semantic voice activity detection. Knowing when a user has stopped speaking is not the same as knowing when they have finished a thought. Semantic VAD is the layer that enables natural turn-taking. It is native to Gradium's STT. Cartesia's Ink is positioned primarily as a transcription model rather than a turn-taking engine.
Streaming STT. Transcription should be available incrementally over WebSocket. Both vendors support this.

Voice Cloning

Clone fidelity, accent and style preservation. Sample audio requirements, clone fidelity, and accent or speaking style handling vary between providers.

Platform

Integrated stack. A production voice pipeline benefits from TTS, STT, and VAD in a single integrated provider.
Enterprise and deployment readiness. Private cloud, on-premise, on-device options, zero data retention, SLA commitments, and concurrency guarantees matter at production scale.

Switching from Cartesia. Gradium's WebSocket TTS accepts the same streaming pattern: open a connection, send text, receive audio chunks. If you are already using Cartesia's WebSocket TTS, switching to Gradium requires updating the endpoint, voice ID, and authentication. The streaming flow is the same, and the json_config parameter gives you additional control over pronunciation, speed, and expressiveness that you can tune after migration.

What Are the Key Differences Between Gradium and Cartesia?

How Does Gradium's TTS Differ from Cartesia for Voice Agents?

Most TTS models are trained and tuned for clean prose: audiobooks, narration, read-aloud content. Voice agents rarely speak in clean prose. They have to say phone numbers, dates, times, URLs, email addresses, order IDs, confirmation codes, street addresses, dollar amounts, and named entities, and they have to say them correctly on the first attempt. Agents that mispronounce a confirmation code or skip a digit in a phone number break the user's trust in a single turn.

Gradium's TTS is tuned specifically for these cases. Phone numbers are grouped and intoned naturally instead of read as a flat digit string. URLs and email addresses are spelled out with correct handling of domains, dots, dashes, slashes, and special characters. Dates and times are pronounced in the regional convention of the target language. Complex named entities (company names, product names, abbreviations) are pronounced consistently across a session.

On latency, both platforms are well positioned for real-time voice agents. The practical threshold for natural-feeling turn-taking is a TTFA under 300 ms delivered over a streaming API, and both Gradium and Cartesia are comfortably within that envelope. Gradium publishes a full TTFA benchmark with measured P50 of 258 ms end-to-end and 214 ms excluding connection establishment, alongside comparisons on the same methodology to the ElevenLabs, Mistral Voxtral and OpenAI models current in March 2026. Those are round-trip figures taken before the Coval methodology change of June 3, 2026. On the Coval perceived-latency read of September 8, 2026, Gradium TTS recorded 214 ms median and Cartesia Sonic 3.6 440 ms, with Sonic 3.5 at 269 ms; Cartesia's own sub-90 ms figure is model inference time excluding network. Where Gradium differentiates is pronunciation of voice-agent content: 81.0% on a 500-sentence hard-case set in August 2026 against 75.1% for Sonic 3.6. How Coval measures perceived time to first audio, and the June 3, 2026 change that introduced the metric, is explained in How a benchmark change produced a faster TTS model.

How Does Gradium's STT with Semantic VAD Compare to Cartesia's Ink?

The most important differentiator in a real-time voice pipeline is not how fast the TTS speaks, but how accurately the system knows when the user has finished speaking. Semantic VAD determines when a speaker has finished a complete thought, not just gone silent. Without it, voice agents fall back on silence thresholds, which produce premature cut-offs or unnatural pauses.

Gradium's STT ships semantic VAD natively. Cartesia's Ink, by contrast, is positioned primarily as a transcription model, leaving turn-taking to the surrounding agent pipeline.

How Does Voice Cloning Compare Between Gradium and Cartesia?

Gradium's voice cloning preserves the accent of the reference speaker from a single 10-second sample, rather than defaulting to a single standard pronunciation. Coverage spans the major regional accents of each supported language. In English, that includes American, British (RP and regional), Australian, Indian, Irish, Scottish, and South African. In French, it covers Metropolitan French, Quebecois, Belgian, Swiss French, and African French. In Spanish, Castilian, Mexican, Argentinian (including Rioplatense pronunciation), Colombian, and Caribbean. In German, High German, Austrian, Swiss German, and Bavarian. In Portuguese, European and Brazilian.

The same cloning pipeline captures speaking style as well as accent: conversational, narrative or audiobook, broadcast, customer-service, expressive and emotional, and whispered delivery. Whatever is in the 10-second sample is what the cloned voice will reproduce. Read more about Instant vs Pro Voice Cloning in Gradium.

How Do Gradium and Cartesia Compare on Language Support?

Cartesia supports 40+ languages. Gradium supports five with native fluency: English, French, Spanish, German, and Portuguese. Gradium adds mid-sentence code-switching across all five, with no latency penalty and no quality degradation. A speaker can shift language within a single sentence and Gradium handles it without a quality drop.

For broad multilingual coverage across 40+ languages, Cartesia is the stronger fit. For deeper handling of the five languages Gradium supports, Gradium is the stronger fit.

How Do Deployment Options Compare?

Gradium is available across the full range of deployment surfaces production teams need. Cloud marketplace for standard SaaS consumption. Private cloud for tenant-isolated deployments inside a customer's account. On-premise for teams with strict data-residency or air-gapped requirements. On-device for latency-critical or offline use cases where the model has to run locally. The same model and API surface is available across all four, so a pilot that ships on the cloud can move to on-prem or on-device without re-architecting the pipeline. Cartesia is certified SOC 2 Type II, HIPAA, and PCI Level 1, and offers enterprise deployment options for regulated environments. Gradium's differentiator is the breadth of explicitly supported surfaces, in particular on-device, delivered from a single model and API.

Gradium TTS: Streaming Text-to-Speech for Real-Time Applications

Gradium's TTS is built for streaming delivery. It connects via WebSocket and supports bidirectional communication, which is the architecture required for voice agents that must speak while staying ready for the next user input.

Capabilities:

Real-time streaming via WebSocket. Audio is delivered incrementally as it is generated, not after the full sentence is complete.
Expressive speech with robust pronunciation. Designed to handle phone numbers, URLs, email addresses, dates, times, and named entities, the inputs that break most TTS models in agent pipelines.
Published TTFA benchmark. P50 258 ms end-to-end, P50 214 ms excluding connection establishment, with full methodology.
High-precision word-level timestamps. Useful for subtitling, lipsync, and interactive transcript display.
Multiple output formats for different integration surfaces.
Advanced configuration via json_config for controlling speed, expressiveness, voice similarity, and text normalization.

The TTS API is available via Python SDK, Rust SDK, and direct WebSocket integration, and is compatible with LiveKit and Pipecat.

Gradium STT: Speech-to-Text with Semantic VAD

Gradium's STT does more than transcribe. Its core differentiator for real-time use cases is semantic VAD: a mechanism that determines when a speaker has finished a thought, not just stopped making sound.

This matters in conversational AI. A standard VAD cuts off after a silence threshold, so the system either interrupts mid-sentence or waits too long. Semantic VAD understands the intent of the utterance and triggers turn-taking at the right moment, producing human-like responsiveness.

Capabilities:

Best-in-class accuracy with controllable latency
Robust performance in noisy environments, designed for real-world deployment
Semantic VAD for smart turn-taking
Streaming via WebSocket

Voice Cloning: Instant and Pro

Gradium offers two voice cloning tiers.

Instant Voice Clone. Create a custom voice from as little as 10 seconds of audio. The clone is immediately available for TTS generation via the API. All paid plans include up to 1,000 Instant Voice Clones per month.

Pro Voice Clone. A fine-tuned model trained on more audio, designed to be indistinguishable from the original speaker. Gradium positions Pro Voice Clones as the highest speaker-similarity option on the market. Pro clones are available from the M plan ($340/month, 5 included) and L plan ($1,615/month, 20 included).

Both clone tiers preserve the accent and speaking style of the reference sample within each of Gradium's five supported languages (see the key differences section above for the full accent and style list).

Both clone types are accessible via the REST API, the Python SDK, or Gradium Studio.

Languages: Native Fluency Across Five Languages

Gradium supports five languages with native fluency: English, French, Spanish, German, and Portuguese.

Mid-sentence code-switching is supported across all five, with no latency penalty and no quality degradation. This is relevant for multilingual AI companions, international customer support agents, and language learning applications.

For deployments that require coverage across a broader set of languages, Cartesia supports 40+ languages and is the stronger fit.

Deployment Options

Gradium supports four deployment surfaces from the same model and API:

Cloud marketplace. Fastest path to production with standard SaaS consumption.
Private cloud. Tenant-isolated deployment inside the customer's cloud account.
On-premise. For teams with strict data-residency, regulatory, or air-gapped requirements (healthcare, financial services, defense).
On-device. For latency-critical or offline use cases where the model runs locally on the end-user device.

Enterprise plans also include zero data retention and SLA commitments. Cartesia is certified SOC 2 Type II, HIPAA, and PCI Level 1, and offers enterprise deployment options for regulated environments. Gradium's differentiator is the breadth of explicitly supported surfaces, in particular on-device, delivered from a single model and API.

Gradium Pricing

Gradium offers a free tier and credit-based paid plans starting at $13/month. For full pricing details, see the Gradium pricing page.

Gradium also runs a Startup Program: seed-funded startups can apply for $2,000+ in free credits, 6 months of full API access, direct engineering support, and early model access.

Who Should Choose Gradium Over Cartesia?

Choose Gradium if you are:

TTS-driven.

Shipping a voice agent that needs robust pronunciation on complex inputs (phone numbers, URLs, email addresses, dates, times, street addresses, order IDs, named entities) across all major languages it supports (English, French, Spanish, German, Portuguese), not just on clean narration
Evaluating TTS with a published, reproducible latency benchmark you can replicate in your own environment

STT-driven.

Building conversational AI agents where natural turn-taking is the key quality driver, and semantic VAD is a hard requirement

Voice-cloning-driven.

Cloning voices that need to preserve a specific accent or speaking style within a language, rather than defaulting to a single standard pronunciation

Platform-driven.

Working in Python or Rust and want officially supported SDKs
Integrating with LiveKit or Pipecat and want a voice layer that connects natively
Deploying across multiple surfaces (cloud marketplace, private cloud, on-premise, on-device) from a single model and API
Requiring enterprise-grade data control, with zero data retention and SLA commitments on enterprise plans
A seed-funded startup looking for production-grade voice AI with generous onboarding credits

Also comparing ElevenLabs or Deepgram? See our dedicated comparison pages.

Cartesia remains the stronger choice if:

Voice preference is the primary requirement: Sonic 3.6 ranked first of 92 at 1,282 Elo on the Artificial Analysis provider-voice board read September 8, 2026. Note that Cartesia's reported sub-90 ms figure is model inference time excluding network, and that on the Coval read of the same date Sonic 3.6 measured 440 ms median perceived time to first audio
You need coverage across 40+ languages

Glossary

State Space Model (SSM). The architecture family behind Cartesia's Sonic line. Cartesia claims lower time to first audio, a lower real-time factor and higher throughput from it.

Delayed Streams Modeling (DSM). The streaming architecture Gradium's models are built on, described in arXiv:2509.08753 (Kyutai).

Perceived time to first audio. Round-trip time to first audio plus the leading silence inside the stream. The Coval headline metric since June 3, 2026.

Hard-case pass rate. The share of adversarial sentences where a native-speaker rater hears every element pronounced correctly and completely.

Credit pricing. Billing against a monthly allocation rather than per character, which requires converting to a per-character rate before providers can be compared.

Part of the provider alternatives cluster, hub at best ElevenLabs alternatives 2026. Its siblings:

Best ElevenLabs alternatives 2026: the cluster hub, the full field ranked on voice quality and price.
ElevenLabs alternative: Gradium: the migration view for teams already on ElevenLabs.
Deepgram alternative: Gradium: the migration view for teams already on Deepgram.
Azure TTS alternative: Gradium: the migration view for teams on Azure AI Speech.
Gradium vs ElevenLabs for voice agents: the measurement view rather than the migration view.
How to compare TTS pricing across providers: the credit maths behind every price in this cluster.

Beyond this topic

A migration is a measurement problem before it is an engineering one. TTS latency benchmark 2026 and TTS WER benchmark 2026 cover how the numbers here are produced, and how to choose a TTS API covers the criteria to rank before you switch.

Gradium's free tier includes 45,000 credits per month with no credit card. API reference and quickstarts are at docs.gradium.ai. For enterprise evaluations, use our contact form.