Cartesia

Cartesia is a San Francisco-based AI company focused on real-time voice synthesis, speech recognition, and state space model (SSM) research.^[1][46][50] Founded in 2023 by researchers from Stanford University's AI Lab, the company builds text-to-speech and speech-to-text systems optimized for interactive applications where latency is the primary constraint: voice agents, customer service automation, accessibility tools, and live audio pipelines.^[1][19] Its flagship product is the Sonic text-to-speech model family. In May 2026 Sonic 3.5 took first place on the Artificial Analysis Speech Arena leaderboard, and its successor Sonic 3.6, released on August 27, 2026, again took first place on the Artificial Analysis leaderboards, according to the company; Cartesia also says its Ink 2 speech-to-text model ranked first on Artificial Analysis's streaming leaderboard, and it markets itself as "the only one ranked #1 on both speech and transcription."^{[31][34][38][39]}

The company sits at an unusual intersection: it is simultaneously a model research lab publishing work on SSM architectures, including the Mamba line, and a commercial voice platform serving developers at API scale.^[13][57] Sonic 3, released in October 2025, generates speech with 90ms model latency, supports 42 languages, and adds laughter and emotion controls; Sonic 3.6 extends language support to 44 languages while replying in under 90ms.^[27][33][38] A parallel speech-to-text line called Ink launched in June 2025, completing Cartesia's ambition to own both sides of the voice pipeline, and a managed voice agent product (first launched as Line in August 2025) sits on top of both models.^[22][23][45] Total funding reached $191 million by late 2025, with Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA among the investors; none of the funding announcements disclosed a valuation.^[2][21][50] As of October 2026 the company says it works from offices in San Francisco, London, and Bangalore.^[46]

Who founded Cartesia and when?

Cartesia traces its origins to the Stanford AI Lab, where a group of PhD students and their advisor spent years developing state-space model architectures as an alternative to the transformer.^[1][19] The company's about page says its founding team "met as PhDs at Stanford AI Lab," and the company lists four co-founders: Karan Goel (CEO and founder), Albert Gu (chief scientist and co-founder), Arjun Desai, and Brandon Yang.^[1] Fortune reported in March 2025 that Stanford professor Christopher Ré was "also on the founding team," and Dell Technologies Capital, an investor, names Ré as a co-founder alongside the other four.^[2][19] Goel told Fortune that all of his co-founders had been Ré's PhD students.^[2] The name comes from René Descartes; Goel said "Cartesian coordinates are very core to mathematics."^[2]

The key theoretical work arrived in several waves. Gu, Goel, and Ré co-authored the S4 (Structured State Space sequence model) paper, posted in October 2021 and given an outstanding paper honorable mention at ICLR 2022, which made the classical state space model practical as a deep sequence model for very long sequences.^[52] In February 2022, Goel, Gu, Chris Donahue, and Ré applied S4 to raw audio in "It's Raw! Audio Generation with State-Space Models," introducing the SaShiMi architecture for waveform generation, an early pointer toward Cartesia's later focus on speech.^[53] Gu completed his Stanford PhD in 2023 with a thesis titled "Modeling sequences with structured state spaces."^[28][51] In December 2023, Gu and Tri Dao (associated with the Flash Attention work) posted the Mamba paper, which introduced a selective state space mechanism; on language modeling the authors reported that their Mamba-3B model outperformed transformers of the same size and matched transformers twice its size, with five times higher inference throughput than transformers.^[13] Mamba demonstrated linear scaling in sequence length versus the quadratic scaling of standard attention, a difference that matters substantially for streaming audio, where a 30-second clip at common codec rates represents thousands of tokens.^[13]

Christopher Ré is a professor of computer science in the Stanford AI Lab and received a MacArthur Fellowship in 2015.^[29][30] Index Ventures describes his Stanford lab as "a hotbed of research and multiple billion-dollar startups," including SambaNova, Snorkel AI, and Together AI.^[4] Goel grew up in New Delhi in a family that ran a scientific equipment manufacturing business, earned a master's degree at Carnegie Mellon, and completed his PhD at Stanford, where he met his co-founders; in 2023 they spun their laboratory research out into Cartesia.^[19]

Cartesia was founded in 2023, emerging from Stanford's AI Lab.^[2][19] The founding thesis was direct: SSMs offered memory and latency characteristics that transformers could not easily replicate for real-time inference, and audio was the natural first application because conversational voice requires fast, continuous generation rather than batch processing.^[19][6] Generating the next frame of audio from an SSM requires updating a fixed-size hidden state rather than attending over the full context window, which means the time per generation step stays constant regardless of how long the conversation has been running.

Albert Gu serves as Cartesia's chief scientist while also leading the Goomba Lab in the Machine Learning Department at Carnegie Mellon University.^[1][57][60] Karan Goel became CEO.^[1] The company launched its first product, Sonic, on May 31, 2024.^[6]

How much funding has Cartesia raised?

On December 12, 2024, Cartesia announced a $27 million seed round led by Index Ventures, with participation from Lightspeed, Factory, Conviction, General Catalyst, A*, SV Angel, and 90 angel investors.^[36] Index Ventures described the same announcement as $22 million in new funding that brought total capital raised to $27 million, and listed angels including the founders of Airtable, Cohere, Datadog, Hugging Face, Mistral, and Zapier.^[4][25] The seed announcement came about six months after Sonic's public launch.^[6][36]

On March 11, 2025, the company announced a $64 million Series A led by Kleiner Perkins.^[2][3] Fortune reported that Index Ventures, Lightspeed, A*, Factory, and Greycroft participated, along with Dell Technologies Capital and Samsung Ventures.^[2] The announcement coincided with the launch of Sonic 2.0 and Sonic Turbo.^[3][26] Total capital raised at that point reached $91 million.^[2]

In late October 2025, Cartesia announced a further $100 million from Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA, alongside the release of Sonic 3.^[21][27] The research firm Sacra describes this round as a "$100M Series A-II" that brought total capital raised to $191 million; the company's own description of the round's series label was not found in the sources reviewed.^[50] None of the announcements reviewed for the seed, the Series A, or the October 2025 round disclosed a valuation.^{[2][4][21][36]}

The company's investors span venture firms (Index, Kleiner Perkins, Lightspeed) and corporate investors (NVIDIA, Dell Technologies Capital, Samsung Ventures).^[2][21] NVIDIA is named among the investors in the October 2025 round.^[21][27]

The Sonic model family

Sonic (original, 2024)

The original Sonic model launched on May 31, 2024, as Cartesia's first commercial TTS product, with a web playground, a low-latency API, a voice library, instant cloning, and voice design controls for speed and emotion.^[6] The company described a model latency of 135ms, "the fastest for a model of this class."^[6] Cartesia said it had built its own state space model inference stack to serve Sonic, and that in its experiments a parameter-matched Cartesia model trained on Multilingual LibriSpeech achieved 20% lower validation perplexity, a 2x lower word error rate, and 1.5x lower time-to-first-audio than widely used transformer implementations.^[6] In October 2024 the company added an alpha release of eight more languages (Hindi, Italian, Korean, Dutch, Polish, Russian, Swedish, and Turkish) to Sonic Multilingual,^[65] and Sonic's last snapshots supported 15 languages.^[37] The streaming API design is a direct consequence of the SSM architecture, which can produce audio frames one at a time without needing to re-read the full prior context on each step.

Sonic 2.0 and Sonic Turbo (March 2025)

Sonic 2.0 launched on March 11, 2025, alongside the Series A announcement.^[3] Cartesia said the model is built on a new state space model architecture and is "twice as large as Sonic, yet runs faster," at 90ms latency for the full model and 40ms for Sonic Turbo, a smaller and faster sibling released at the same time.^[3][26] In blind head-to-head evaluations on 100 held-out voices, Cartesia said 1.5 times as many people preferred Sonic 2.0 over the next best provider.^[3]

The first Sonic 2 and Sonic Turbo snapshots (dated March 7, 2025) supported 15 languages: English, French, German, Spanish, Portuguese, Chinese, Japanese, Hindi, Italian, Korean, Dutch, Polish, Russian, Swedish, and Turkish.^[37] The release also introduced two new API endpoints: a voice changer (applying a voice style to existing audio) and infill editing (replacing a segment of audio while preserving surrounding context).^[3] Cartesia's documentation set October 20, 2026 as the date after which sonic-2 and sonic-turbo and all of their snapshots would stop working, with customers directed to migrate to Sonic 3.6.^[37]

Sonic 3 (October 2025)

Sonic 3 launched in late October 2025 alongside Cartesia's $100 million funding round; its first API snapshot is dated October 27, 2025.^[21][27][37] Coverage of the launch reported a model latency of 90 milliseconds, an end-to-end response time of 190 milliseconds, and support for 42 languages.^[21][27]

Languages added in Sonic 3 include Arabic, Bulgarian, Croatian, Czech, Danish, Finnish, Greek, Hebrew, Hungarian, Indonesian, Romanian, Slovak, Tagalog, Ukrainian, Vietnamese, and several Indian languages such as Tamil, Telugu, Bengali, and Marathi.^[37]

The most distinctive feature of Sonic 3 is emotional expression. Amazon Web Services described the model as providing "advanced controllability through API parameters and SSML tags for volume, speed, and emotion adjustments," with "natural laughter support," stable voices for voice agents, and emotive voices for expressive characters.^[33]

Retell AI made Sonic 3 available on its platform with support for 27 new languages, custom pronunciations, and speed and volume controls.^[9] In February 2026, Amazon Web Services made Sonic 3 available through Amazon SageMaker JumpStart, describing it as delivering "high naturalness, accurate transcript following, and industry-leading latency" with sub-100ms latency and 42-language support, so enterprises can deploy the model inside their own AWS accounts.^[33]

Sonic 3.5 (May 2026)

Cartesia's documentation dates the first Sonic 3.5 snapshot to May 4, 2026, and describes it as a model with sub-90ms latency and native support for 42 languages that voices confirmation codes, phone numbers, and heteronyms correctly without preprocessing.^[43] On May 22, 2026, Artificial Analysis announced that Sonic 3.5 had taken first place on its Speech Arena leaderboard, ahead of Inworld Realtime TTS 1.5 Max and Google's Gemini 3.1 Flash TTS, with an Elo score of 1,218 based on 1,144 arena appearances.^[34] Artificial Analysis listed the model's price at $39 per million characters and noted 42 languages, including nine Indian languages, and more than 500 voices.^[34] CEO Karan Goel posted the same day: "Our new speech model Sonic-3.5 is now #1 on Artificial Analysis's leaderboard. Less than 2 years ago, we released Sonic-1, the fastest speech model in the world. Sonic-3.5 now brings the best speech model for conversation with the lowest latency in production."^[32] Cartesia later presented Sonic 3.6 and Ink 2 together as a combined voice stack on a dedicated launch page.^[31]

Sonic 3.6 (August 2026)

Sonic 3.6 was released on August 27, 2026, a little under four months after the first Sonic 3.5 snapshot (Cartesia's announcement described it as landing "just two months after Sonic-3.5").^[37][38] Cartesia said the model was the result of rebuilding "our data, model architecture, training and evaluation from the ground up," and that it again took first place on the Artificial Analysis leaderboard on both the controlled-voice and provider-voice boards, displacing Sonic 3.5 on the controlled board.^[38] In the company's own blind tests across fifteen locales, listeners preferred Sonic 3.6 over Sonic 3.5 in up to 93% of comparisons; against ElevenLabs' Eleven v3, native-speaker panels chose Sonic 3.6 92% of the time in US English and 67% of the time in Brazilian Portuguese.^[38]

Sonic 3.6 supports 44 languages, adding Odia and Urdu, and the company counts 61 locales (11 of them Indic) and more than 500 preset voices.^[7][38] It adds Hinglish code-switching between Hindi and English within one generation, a locale field for reading dates, times, and numbers the way a local speaker would, and better accent retention for instant voice clones.^[7][38] Cartesia says Sonic 3.6 replies in under 90ms, generates 132 characters per second, and can be deployed in a customer's own cloud through Baseten, Amazon SageMaker, or Together AI.^[38]

Ink: speech-to-text

On June 10, 2025, Cartesia released Ink, a family of streaming speech-to-text models designed to complement Sonic in full-stack voice pipelines.^[22] The debut model, Ink-Whisper, is a variant of OpenAI's Whisper (specifically whisper-large-v3-turbo) rearchitected for low-latency transcription in conversational settings.^[22]

Ink-Whisper addresses the specific failure modes that matter in voice agent deployments: telephony audio artifacts, proper nouns and domain-specific terminology, background noise, disfluencies and silence, and accent variation.^[22] The model uses dynamic chunking to handle variable-length audio chunks that end at semantically meaningful points, rather than the full 30-second chunks standard Whisper prefers.^[22]

Performance claims center on time-to-complete-transcript (TTCT), the time until the full transcript is ready after the user stops talking; Cartesia reported a median TTCT of 66ms and a P90 of 98ms, the fastest of the streaming models it tested at launch.^[22] At launch it was priced at 1 credit per second of audio, or $0.13 per hour on the Scale plan.^[22] Integrations with Vapi, LiveKit, and Pipecat were available at launch.^[22]

The line advanced with Ink 2. Goel announced on May 28, 2026, that Ink-2 was out and ranked first on Artificial Analysis, citing "low latency, fast eager mode and built in semantic endpoints to detect when users are done talking."^[35] Cartesia's launch post, dated July 9, 2026, says Ink 2 ranked first on Artificial Analysis's streaming leaderboard for lowest word error rate as of June 2026, and reports a time-to-final-transcript of 0.1 seconds.^[39] The model emits turn events natively (turn.start, turn.eager_end, and turn.end), so an agent can begin generating a reply before a turn is confirmed complete, without a separate voice activity detector.^[39] On the AppTek multi-accent call-center benchmark, Cartesia reported 8% word error rate for Ink 2 against 10% for Deepgram Flux and 12% for ElevenLabs Scribe v2 Realtime.^[39] In August 2026 the company added keyterm prompting (up to 100 keyterms per connection) and configurable turn detection.^[40] Ink 2 initially launched as an English model; since September 17, 2026, it also transcribes French, Hindi, Japanese, and Spanish.^[39][41]

With Ink handling the speech-to-text step and Sonic handling text-to-speech, Cartesia's agent platform (described below) can offer a fully Cartesia-owned voice pipeline rather than routing through external ASR providers.^[45]

Line and Managed Agents: voice agent platform

Line is Cartesia's voice agent development and deployment platform, announced on August 19, 2025.^[23] Where Sonic and Ink are low-level model APIs, Line is a higher-level platform for building, iterating, and operating voice agents without assembling the infrastructure stack from scratch.^[23]

The platform launched as code-first: agents are written as code using the Line SDK, which can be developed locally and deployed with a single command.^[23] From a text prompt or template, a developer can have a first agent running in minutes.^[23]

Key Line features at launch included:

• Cartesia's Ink STT and Sonic TTS deployed alongside the agents for the fastest latencies^[23]
• Background reasoning while the primary conversation continues^[23]
• Call recording, audio and transcript storage, latency metrics, and logs for auditing and debugging^[23]
• Custom LLM-as-a-judge metrics to rate calls on criteria such as user satisfaction and call success^[23]
• CLI and GitHub integrations for local development workflows, plus fully on-premises deployment for enterprises^[23]

By October 2026 the product was marketed as Managed Agents, a fully managed runtime with built-in evaluations, SIP trunking, and Cartesia-provisioned phone numbers, in which developers configure agents through a playground or API and choose from a catalog of managed LLMs that Cartesia runs and bills as passthrough.^[45][44][15] Cartesia's documentation says it stops hosting code-first Line SDK agents on December 1, 2026, and provides a guide for moving them to Managed Agents; self-hosted agent code is unaffected.^[44] As of October 2026, agent calls are priced at $0.06 per minute, plus $0.014 per minute for a Cartesia-provided phone number.^[15]

Voximplant, a cloud communications platform, announced a native connector for Cartesia Line agents in February 2026, so that Line-built voice agents can handle calls over the public phone network, SIP, WebRTC, and WhatsApp Business Calling.^[24][69]

Why does Cartesia use state-space models for audio?

Cartesia's technical differentiation rests on applying SSMs to audio generation in place of the transformer-based architectures used by most competitors.^[6][36]

The core computational distinction concerns how each architecture handles prior context. A transformer with standard self-attention processes all previous tokens in parallel, with compute scaling as O(n squared) in sequence length. An SSM maintains a fixed-size hidden state updated recurrently, with compute scaling as O(n).^[13] For audio specifically, this difference is significant: a 30-second clip sampled at typical codec rates represents thousands of tokens, and a multi-turn conversation may span hundreds of thousands.

Beyond raw compute efficiency, SSMs have a favorable property for streaming generation. Because they operate recurrently, the model can generate the next audio frame without re-reading the entire prior context, so the computation required per output step stays constant regardless of conversation length.^[13]

In its December 2024 seed announcement, Cartesia said it had created "a new SSM architecture for multi-stream models that continuously reason and generate over multiple data streams of different modalities in parallel," and had trained an end-to-end voice generation model with it that gives fine-grained control over the input text to prevent hallucinations.^[36]

The architecture also benefits from constant memory consumption during inference. Unlike a transformer KV cache that grows with context length, an SSM's hidden state stays the same size regardless of how long the conversation has run.^[57] This matters for infrastructure cost at scale: serving many simultaneous long-running voice calls does not require allocating proportionally larger memory per call.

Cartesia's researchers acknowledge a trade-off. Writing about Mamba-3, the team noted that "there is no free lunch": the same fixed state size that enables efficient computation forces the model to compress all past information into one representation, the opposite of a transformer's growing KV cache.^[57] The Mamba selective state mechanism addressed part of this by letting the model learn which input information to propagate or forget, but it remains an area of ongoing research.^[13][57]

How low is Cartesia's latency?

Cartesia distinguishes model latency (the time between receiving input text and producing the first audio internally) from end-to-end latency, which includes the network round trip that a developer's application actually observes.^[21][27] Its published figures have fallen from 135ms model latency for the original Sonic to 90ms for Sonic 2.0 and Sonic 3, 40ms for Sonic Turbo, and under 90ms for Sonic 3.5 and 3.6.^{[3][6][27][38][43]}

Vapi, after integrating with every major TTS provider, chose Cartesia as its default provider; Cartesia's case study says it is "the only provider with end-to-end latency consistently under 200 ms across all languages."^[8] On Retell's platform, Cartesia says its time-to-first-audio is two to three times faster than the next best provider, with an error rate below 0.1%.^[9] Cartesia's ElevenLabs comparison page, which credits Coval, an independent voice AI evaluation platform, for its P90 latency measurements, shows Sonic 3.5 with a median time-to-first-audio of 320ms and a standard deviation of 62ms in its latency-distribution chart.^[16]

API and pricing

Cartesia exposes three ways to call its text-to-speech models: a bytes endpoint that streams audio as a single HTTP response, server-sent events (SSE), and WebSockets. All three stream audio as it is produced; WebSockets additionally support multiple generations per connection and "continuations" for feeding a transcript in pieces, such as tokens arriving from an LLM.^[66] Output can be raw PCM, WAV, or MP3.^[67]

The API accepts plain text input along with parameters for voice, language or locale, speed, volume, and emotion, and separate endpoints handle voice cloning, voice changing, and infill editing.^[3][7][33]

Pricing uses a credit system. For standard TTS, 1 credit equals 1 character of input text, and one minute of generated audio takes roughly 750 to 800 credits; for Ink speech-to-text, 3 credits equal 1 second of audio.^[15] Plans as listed in October 2026:

Each paid plan also includes prepaid usage for Managed Agents equal to its monthly price (for example, $299 on the Scale plan), and the Free plan includes $1; and overage rates range from $65 per million credits on Pro to $38 per million on Scale.^[15] Cartesia says it is SOC 2 Type II certified and offers on-premises and virtual private cloud deployments.^[15]

For enterprise customers, Cartesia also runs deployments for regulated workloads on Together AI's PCI-compliant environments, which are configured with data isolation, encryption at rest and in transit, and audit logging.^[17] The company says its platform covers SOC 2, PCI, HIPAA, and GDPR requirements; it announced GDPR compliance in September 2025.^[38][47]

How does Cartesia voice cloning work?

Cartesia offers two voice cloning tiers with different use cases.

Instant voice cloning (IVC) creates a custom voice from a short recording in seconds. Cartesia's documentation says 10 seconds of audio is enough to get started, and up to 60 seconds helps retain the speaker's accent on Sonic 3.6 and newer; older models learn only from the first 10 seconds.^[68] IVC is available on the Pro tier and above.^[15]

Professional voice cloning (PVC), introduced on May 6, 2025, fine-tunes Sonic on a larger set of recordings to capture tone, cadence, style, and recording environment.^[10] Cartesia's documentation says a PVC trains on 30 minutes or more of audio from the same speaker.^[68] PVC is designed for brand voices, private voice libraries, and personal AI avatars.^[10] As of October 2026, Cartesia states that training a PVC is free and that speech generated with a PVC costs 1 credit per character, the same as instant cloning; PVC slots are included with the Startup plan (2 slots) and Scale plan (4 slots).^[10][15]

Cloned voices can be localized into other languages and accents; the pricing page lists a cost of 225 credits per added voice accent.^[15] On September 23, 2026, Cartesia introduced Multilingual Voices: more than 50 library voices that speak up to 25 languages natively, plus the option to make a customer's own clone multilingual so that one voice identity carries across markets.^[61] Every language and accent is scored by native speakers on a rubric, according to the company.^[61]

Voice cloning is also accessible through Cartesia's web playground for prototyping without writing code.^[6][10]

Cartesia's terms of service prohibit using the service to make available audio of any person's voice without that person's express permission, or to impersonate any person or entity.^[71] The company does not publish details of technical measures used to detect misuse.

Infrastructure partnership with Together AI

Cartesia runs its production voice serving and model training on Together AI's GPU clusters rather than operating its own data centers.^[17] Together provides the GPU compute, networking, storage, and cluster operations; within each node GPUs connect over NVLink, and inter-node training runs over InfiniBand with GPU-direct RDMA, with WekaFS mounted over RDMA to handle the millions of small random reads typical of audio training.^[17] Cartesia trains on Together's H100 80GB clusters with multi-node jobs scheduled through Slurm and sharded with PyTorch FSDP.^[17]

The arrangement gives Cartesia deep cluster-level access, including SSH and direct file system access, to run a custom inference engine optimized specifically for SSM architectures rather than relying on generic serving stacks designed for transformer inference.^[17] "Because our voice models use state space architectures, we need to run a purpose-built inference engine," said Kunal Shah, Cartesia's head of platform, in Together's case study.^[17]

Cartesia says the relationship has been running in production since Sonic's launch, powering millions of audio minutes daily on Together's GPU clusters.^[17][18] Cartesia is also a dedicated model partner on Together AI's Voice Platform, where Sonic is offered as a native endpoint for Together's enterprise customers.^[18]

Who uses Cartesia?

Cartesia's go-to-market relies primarily on developer-led adoption through its API, playground, and documentation, with ecosystem integrations into voice agent platforms as a major distribution channel.^[23][50] Kleiner Perkins and Fortune reported in March 2025 that more than 10,000 customers used Sonic.^[2][5] In August 2025 the company said "tens of thousands of developers" had built voice agents on Sonic and Ink, and coverage of the October 2025 round said thousands of companies, including ServiceNow, Cresta, and Decagon, used Sonic for millions of voice interactions each month.^[23][27]

Vapi

Vapi is a voice agent orchestration platform that handles turn detection, LLM routing, and TTS provider connections for developers building phone-based AI agents. One of Cartesia's earliest design partners, Vapi chose Cartesia as its default model provider after integrating with every major TTS provider, and used it to power its homepage demo.^[8] Vapi cited consistent sub-200ms end-to-end latency across all supported languages as a deciding factor.^[8] Ink-Whisper was subsequently added as an STT option within Vapi as well.^[22]

Retell AI

Retell AI is a competing voice agent platform that also integrated Cartesia as a primary voice provider and automatic failover.^[9] Retell's customers add Cartesia voices by flipping a switch in the platform, without re-engineering; Cartesia describes the adoption as the fastest Retell had seen from a new provider.^[9]

Together AI

Together AI operates as both an infrastructure partner (described above) and a distribution channel. Enterprises that use Together AI's model serving platform can access Sonic through Together AI's APIs and billing rather than directly through Cartesia.^[18]

ServiceNow and enterprise customers

Cartesia's voice technology is integrated with ServiceNow AI Voice Agents, part of ServiceNow's AI Experience; for ServiceNow customers Cartesia offers its models, inference engine, and orchestration with fully air-gapped on-premises deployment.^[49]

India

India has become a focus market. Cartesia lists Indian voice AI customers including Fundamento (lending and insurance) and Bolna, and a partnership with Blue Machines AI to deliver India-resident, multilingual voice agents for regulated sectors such as banking, financial services and insurance (BFSI) and healthcare.^[63][64] On October 12, 2026, Cartesia announced that Raise Financial Services, which runs the Dhan trading platform, uses its models for part of Dhan's voice bot interactions, including conversational KYC; the companies said more than 75% of those interactions happen in languages other than English, and Cartesia said Sonic supports 11 Indic languages with code-switching.^[62]

Other platforms and customers

Cartesia's customer page features EliseAI, which it says runs more than 5 million voice calls a month on Cartesia, and Replicant, which runs millions of enterprise contact-center conversations on Sonic.^[48][64] Other listed customers and partners include Quora's Poe, LiveKit, Daily, Tavus, Captions, Rasa, 11x, Assort Health, and Maven AGI.^[64]

How does Cartesia compare with ElevenLabs?

ElevenLabs and Cartesia are frequently compared in developer discussions of voice AI APIs, and Cartesia maintains a comparison page aimed at ElevenLabs customers.^[16] The claims on that page are the company's own marketing and should be read as such.

According to that page, Sonic 3.6 scored 1,282 Elo on the Artificial Analysis Provider Voice Arena as of August 2026, against 1,177 for Eleven v3, 1,103 for Turbo v2.5, and 1,084 for Flash v2.5; Cartesia's native-speaker panels chose Sonic 3.6 over Eleven v3 in all seven locales it tested.^[16] Cartesia also says it supports 44 languages from one model while ElevenLabs offers 32 languages on its real-time models, that ElevenLabs' real-time models respond in about 250ms, and that Cartesia is about 50% cheaper.^[16] ElevenLabs offers its own speech-to-text model (Scribe) and agent platform (ElevenAgents).^[16][39] In the per-dataset word error rates Cartesia published for June 2026, ElevenLabs Scribe v2 Realtime had lower error rates than Ink 2 on the AA-AgentTalk and VoxPopuli datasets, while Ink 2 was lowest on Earnings22.^[39]

For voice agents where every additional fraction of a second is audible to the caller, latency is the main reason developers cite for choosing Cartesia, as in Vapi's and Retell's case studies.^[8][9]

On-device deployment

On August 28, 2024, Cartesia open-sourced Edge, an Apache 2.0 library for running SSMs on device hardware, starting with Apple's M-series chips through custom Metal kernels for Mamba-2.^[11] The same release introduced Rene, an open 1.3B-parameter hybrid language model with a Mamba-2 backbone trained on 1.5 trillion tokens of the Dolma-1.7 dataset, and a private beta of Sonic On-Device, which Cartesia described as the first ultra-realistic generative voice model to support low-latency real-time streaming on device.^[11] Edge itself hosts open-weights SSM language models; Sonic On-Device was offered separately as a private beta.^[11]

On-device deployment addresses data privacy and connectivity requirements: Cartesia argued that on-device models avoid network latency, work without connectivity, and keep data on the hardware.^[11] As of October 2026 the company markets the same models and agents for cloud, on-premises, and on-device deployment.^[48]

The Llamba model family, described in a February 2025 technical report, extends this work to language modeling.^[12][58] Llamba-1B, Llamba-3B, and Llamba-8B are recurrent language models distilled from the Llama 3 series into a Mamba-2-based architecture using the MOHAWK distillation method, with less than 0.1% of the training data typically used for models of similar size.^[58] On an NVIDIA H100, Cartesia reported that Llamba-8B reached up to 12 times the throughput of Llama-3.1-8B at a generation length of 8,192 tokens, and the models were released for Edge with an implementation for Apple silicon.^[12]

Research contributions

Beyond Sonic and Ink, Cartesia publishes SSM research, much of it in collaboration with Gu's Goomba Lab at Carnegie Mellon and with Tri Dao at Princeton.^[56][57]

The Mamba-3B-SlimPJ release in December 2023, made in partnership with Cartesia and Together, showed SSMs matching the best transformer architectures at the 3B parameter scale on language modeling benchmarks, an early proof point that SSMs are not limited to specialized audio tasks.^[59]

Llamba (February 2025) showed that distilling transformer knowledge into SSM architectures works at scale, with higher inference throughput and larger batch sizes than transformer models at comparable benchmark performance.^[58]

In July 2025 Cartesia announced a research collaboration on hierarchical networks (H-Nets), described in the paper "Dynamic Chunking for End-to-End Hierarchical Sequence Modeling" by Sukjun Hwang, Brandon Wang, and Albert Gu.^[55] H-Net learns to segment raw bytes into chunks jointly with the rest of the model, and the authors reported that a byte-level H-Net outperformed a strong BPE-tokenized transformer when compute- and data-matched, with nearly 4x better data efficiency on DNA sequences.^[55] Cartesia released H-Net checkpoints on Hugging Face and described hierarchical modeling as the "key next step" after SSMs for multimodal and long-context models.^[70]

Mamba-3, posted to arXiv on March 16, 2026, and accepted at ICLR 2026, was written by researchers at Carnegie Mellon University, Princeton University, Together AI, and Cartesia.^[14][56] The paper introduced three improvements to the Mamba-2 layer: exponential-trapezoidal discretization, complex-valued state updates, and a multi-input, multi-output (MIMO) recurrence formulation.^[56] At the 1.5B parameter scale, Mamba-3 improved average downstream accuracy by 0.6 percentage points over the next best model (Gated DeltaNet), and the MIMO variant added another 1.2 points for a total gain of 1.8 points without increasing decoding latency; across state-size experiments, Mamba-3 matched Mamba-2's perplexity with half the state size.^[56] The Mamba-3 team, in a post cross-posted from Gu's Goomba Lab, framed the work as an "inference-first" design for workloads dominated by post-training, deployment, and real-time use.^[57]

What is Cartesia used for?

Voice agents and telephony

The primary commercial application for Cartesia's technology is automated voice agents: AI systems handling inbound and outbound phone calls for appointment scheduling, customer service, sales qualification, recruiting interviews, and similar tasks.^[23][39][64] Companies like Vapi, Retell, and Thoughtly build orchestration layers on top of Cartesia's models, while Cartesia's own Managed Agents product competes in the same layer.^[8][9][45]

Latency is the variable that determines whether these interactions feel natural or mechanical. A cascaded voice pipeline runs speech recognition, LLM inference, and TTS in sequence, so each step's contribution adds to the delay a caller hears. Cartesia argues that transcription latency, like synthesis latency, directly affects how natural a conversation feels, which is why it publishes time-to-first-audio and time-to-final-transcript figures for its models.^[22][39]

Content creation and media production

Cartesia's voice cloning and voiceover features serve creators who need consistent narration across large volumes of output. In its Series A announcement the company said it had "helped tens of thousands of creators make their content more accessible," and its listed customers include Captions.^[3][64]

Localization

With 44-language support in Sonic 3.6 and voice cloning that preserves vocal identity across languages, Cartesia is used in localization workflows where content originally recorded in one language needs to be converted to another while retaining the speaker's voice characteristics.^[38][61] Its Multilingual Voices release in September 2026 lets a single brand voice speak up to 25 languages natively.^[61]

Timeline

Limitations

Several constraints are relevant for teams evaluating Cartesia.

Sonic 3.6 covers 44 languages, including many Indian languages, but no sub-Saharan African language such as Swahili, Yoruba, or Amharic appears on its supported-language list.^[7] Ink 2, the company's flagship speech-to-text model, supported five languages (English, French, Hindi, Japanese, and Spanish) as of October 2026, while the older Ink-Whisper model covers a much longer list.^[41][42]

Cartesia retires older models on fixed schedules. Its documentation set October 20, 2026 as the cutoff after which sonic-2, sonic-turbo, and the October 2025 sonic-3 snapshot would stop working, so customers on those models had to migrate.^[37] Code-first agents hosted on the Line SDK must move to Managed Agents before December 1, 2026, unless they are self-hosted.^[44]

Professional voice clones are limited by plan: Free and Pro plans include none, Startup includes two slots, and Scale four.^[15]

Most of the head-to-head quality and latency comparisons Cartesia publishes come from its own evaluations or its marketing pages. Independent rankings such as Artificial Analysis's arenas are a more neutral reference, but they shift as new models are released; Sonic 3.6 displaced Sonic 3.5 about three months after Sonic 3.5 reached first place.^[16][34][38]

Continue reading

• MMambaMamba is a neural network architecture for sequence modeling that uses selective state space models (SSMs) to process sequential data in linear time with respect to sequence lengthMentioned 25 times above
• ElevenLabsElevenLabs is a voice and audio artificial intelligence company that builds text-to-speech AI, voice cloning, AI dubbing, generative sound effects, music synthesis, and conversational voice agent technology.Mentioned 11 times above
• VapiVapi is a voice AI orchestration platform that lets software developers build, deploy, and scale AI phone agents through a programmable API.Mentioned 9 times above
• MMamba-3Mamba-3 is a sequence-modeling architecture in the state space model (SSM) family, introduced in March 2026 by researchers at Carnegie Mellon University and Princeton.Mentioned 7 times above
• HH-Net (dynamic chunking)H-Net, short for Hierarchical Network, is a tokenizer-free neural sequence model that learns to segment raw bytes into content-adaptive "chunks" as part of ordinary end-to-end training, rather than relying on…Mentioned 6 times above
• SState space model (deep learning)A state space model (SSM) in deep learning is a class of sequence model that maps an input sequence to an output sequence through a fixed-size latent stateMentioned 4 times above

More from See also

• List of neolabs
• Neolab
• Albert Gu
• Christopher Ré
• Retell AI
• Voice AI
• Text-to-Speech
• Deepgram
• Flash Attention

Topics:AI CompaniesAI ModelsSpeech & Audio AIVoice AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6Cartesia Company page. cartesia.ai/company
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Fortune, "Exclusive: Cartesia, voice AI startup, raises $64 million Series A." March 11, 2025. fortune.com/2025/03/11/exclusive-cartesia-voice-ai-startup-raises-64-million-series-a/
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Cartesia blog, "Series A and the future of voice AI." March 11, 2025. cartesia.ai/blog/series-a
4. ^1 ^2 ^3Index Ventures, "Building the Next Generation of Real-Time AI Models." indexventures.com/perspectives/building-the-next-generation-of-real-time-ai-models-our-investment-in-cartesia/
5. ^Kleiner Perkins, "Cartesia: Pioneering real-time voice AI." kleinerperkins.com/perspectives/cartesia-pioneering-real-time-voice-ai/
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Cartesia blog, "Announcing Sonic: a low-latency voice model for lifelike speech." May 31, 2024. cartesia.ai/blog/sonic
7. ^1 ^2 ^3 ^4Cartesia docs, latest text-to-speech model page (describes Sonic 3.6 as of October 2026). docs.cartesia.ai/build-with-cartesia/tts-models/latest
8. ^1 ^2 ^3 ^4 ^5Cartesia customer story, "Vapi chooses Cartesia as their default provider for voice agents." cartesia.ai/customers/vapi
9. ^1 ^2 ^3 ^4 ^5 ^6Cartesia customer story, "How Cartesia Powers Retell's Voice Agents at Scale." cartesia.ai/customers/retell
10. ^1 ^2 ^3 ^4 ^5Cartesia blog, "Introducing Professional Voice Cloning." May 6, 2025. cartesia.ai/blog/pro-voice-cloning
11. ^1 ^2 ^3 ^4 ^5 ^6Cartesia blog, "The on-device intelligence update." August 28, 2024. cartesia.ai/blog/on-device
12. ^1 ^2Cartesia blog, "Llamba: scaling distilled recurrent models for efficient language processing." March 5, 2025. cartesia.ai/blog/llamba-distillation
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7Gu, Albert and Dao, Tri. "Mamba: Linear-Time Sequence Modeling with Selective State Spaces." arXiv:2312.00752. December 2023.
14. ^1 ^2OpenReview, "Mamba-3: Improved Sequence Modeling using State Space Principles." ICLR 2026. openreview.net/forum?id=HwCvaJOiCj
15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Cartesia pricing page (accessed October 2026). cartesia.ai/pricing
16. ^1 ^2 ^3 ^4 ^5 ^6Cartesia, "ElevenLabs alternative for voice agents" (comparison page, accessed October 2026). cartesia.ai/vs/cartesia-vs-elevenlabs
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7Together AI, "How Cartesia Runs Real-Time Voice AI on Together AI's GPU Infrastructure." together.ai/customers/cartesia
18. ^1 ^2 ^3Cartesia customer story, "Together AI Chooses Cartesia as Dedicated Model Partner for Enterprise Voice AI." cartesia.ai/customers/together-ai
19. ^1 ^2 ^3 ^4 ^5 ^6Dell Technologies Capital, "Inside Cartesia's Jump from Research to Voice AI Leadership." delltechnologiescapital.com/resources/cartesia-voice-ai
20. Cartesia blog, "State of voice AI 2024." December 19, 2024. cartesia.ai/blog/state-of-voice-ai-2024
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Traded VC, "Cartesia Secures $100 Million Investment From Kleiner Perkins Index Ventures Lightspeed And NVIDIA." October 29, 2025. traded.co/vc/deal/cartesia-secures-100-million-investment-from-kleiner-perkins-index-ventures-lightspeed-and-nvidia/
22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Cartesia blog, "Introducing Ink: speech-to-text models for real-time conversation." June 10, 2025. cartesia.ai/blog/introducing-ink-speech-to-text
23. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Cartesia blog, "Introducing Line: The Modern Voice Agent Development Platform." August 19, 2025. cartesia.ai/blog/introducing-line-for-voice-agents
24. ^GlobeNewswire, "Voximplant Brings Cartesia Line Voice Agents into Real Calls." February 12, 2026.
25. ^Fundraise Insider, "Cartesia Secures $27 Million to Revolutionize Real-Time AI with State-Space Models." fundraiseinsider.com/blog/cartesia/
26. ^1 ^2Maginative, "Cartesia Raises $64M to Advance Real-Time Voice AI with Sonic 2.0." maginative.com/article/cartesia-raises-64m-to-advance-real-time-voice-ai-with-sonic-2-0/
27. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9AIM Media House (Deployed), "Cartesia Raises $100 Million to Transform Real-Time Voice AI with Sonic-3". October 28, 2025.
28. ^Albert Gu Google Scholar. scholar.google.com/citations?user=DVCHv1kAAAAJ
29. ^Stanford Profiles, "Christopher Re, Professor of Computer Science". Accessed October 2026.
30. ^MacArthur Foundation, "Christopher Ré, Computer Scientist, Class of 2015".
31. ^1 ^2Cartesia, launch page "Introducing Sonic-3.6 and Ink-2" (accessed October 2026). cartesia.ai/launch
32. ^Karan Goel (krandiash), post on X announcing Sonic-3.5 reaching #1 on the Artificial Analysis leaderboard. May 22, 2026. x.com/krandiash/status/2057881349852463561
33. ^1 ^2 ^3 ^4 ^5Amazon Web Services, "Cartesia Sonic 3 text-to-speech model is now available on Amazon SageMaker JumpStart." February 4, 2026. aws.amazon.com/about-aws/whats-new/2026/02/cartesia-sonic-3-on-sagemaker-jumpstart/
34. ^1 ^2 ^3 ^4 ^5Artificial Analysis, post on X: "Cartesia's Sonic-3.5 takes the #1 spot on the Artificial Analysis Speech Arena Leaderboard." May 22, 2026. x.com/ArtificialAnlys/status/2057878247782908109
35. ^1 ^2Karan Goel (krandiash), post on X announcing Ink-2 at #1 on Artificial Analysis. May 28, 2026. x.com/krandiash/status/2060044523162697775
36. ^1 ^2 ^3 ^4 ^5 ^6Karan Goel, Cartesia blog, "Announcing our seed round". December 12, 2024.
37. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Cartesia docs, "Older Models" (text-to-speech). Accessed October 2026.
38. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Cartesia TTS Research, Cartesia blog, "Introducing Sonic-3.6". August 27, 2026.
39. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Eli Pugh, Cartesia blog, "Introducing Ink-2: The #1-ranked STT built for voice agents". July 9, 2026 (updated after September 17, 2026).
40. ^Lisa Bedin and Eli Pugh, Cartesia blog, "New to Ink-2: Keyterm prompting and configurable turn detection". August 11, 2026.
41. ^1 ^2Cartesia docs, "Ink 2". Accessed October 2026.
42. ^Cartesia docs, "Older Models" (speech-to-text). Accessed October 2026.
43. ^1 ^2Cartesia docs, "Sonic 3.5". Accessed October 2026.
44. ^1 ^2 ^3 ^4Cartesia docs, "Migrate from the Line SDK". Accessed October 2026.
45. ^1 ^2 ^3 ^4Cartesia, "Managed Agents" product page. Accessed October 2026.
46. ^1 ^2Cartesia, "Careers". Accessed October 2026.
47. ^Karan Goel, Cartesia blog, "Cartesia achieves GDPR compliance". September 24, 2025.
48. ^1 ^2Cartesia, homepage. Accessed October 2026.
49. ^Cartesia customer story, "Cartesia Voice AI Technology Integrated with ServiceNow AI Voice Agents".
50. ^1 ^2 ^3 ^4Sacra, "Cartesia" company profile. Accessed October 2026.
51. ^Gu, Albert. "Modeling sequences with structured state spaces". PhD thesis, Stanford University, 2023. Stanford Digital Repository.
52. ^1 ^2Gu, Albert; Goel, Karan; Ré, Christopher. "Efficiently Modeling Long Sequences with Structured State Spaces". arXiv:2111.00396. October 31, 2021 (ICLR 2022).
53. ^Goel, Karan; Gu, Albert; Donahue, Chris; Ré, Christopher. "It's Raw! Audio Generation with State-Space Models". arXiv:2202.09729. February 20, 2022.
54. ^Dao, Tri; Gu, Albert. "Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality". arXiv:2405.21060. May 31, 2024 (ICML 2024).
55. ^1 ^2 ^3 ^4Hwang, Sukjun; Wang, Brandon; Gu, Albert. "Dynamic Chunking for End-to-End Hierarchical Sequence Modeling". arXiv:2507.07955. July 10, 2025.
56. ^1 ^2 ^3 ^4 ^5 ^6Lahoti, Aakash; Li, Kevin Y.; Chen, Berlin; Wang, Caitlin; Bick, Aviv; Kolter, J. Zico; Dao, Tri; Gu, Albert. "Mamba-3: Improved Sequence Modeling using State Space Principles". arXiv:2603.15569. March 16, 2026 (ICLR 2026).
57. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Mamba 3 Team, Cartesia blog, "Mamba-3: An Inference-First State Space Model". March 11, 2026.
58. ^1 ^2 ^3 ^4Bick, Aviv; Katsch, Tobias; Sohoni, Nimit; Desai, Arjun; Gu, Albert. "Llamba: Scaling Distilled Recurrent Models for Efficient Language Processing". arXiv:2502.14458. February 20, 2025.
59. ^1 ^2 ^3Albert Gu and Tri Dao, Cartesia blog, "Mamba-3B-SlimPJ: State-space models rivaling the best Transformer architecture". December 14, 2023.
60. ^Goomba Lab, Carnegie Mellon University, "People". Accessed October 2026.
61. ^1 ^2 ^3 ^4 ^5Akki Shafiqullah and Aiqi Liu, Cartesia blog, "Introducing Multilingual Voices". September 23, 2026.
62. ^Dhwani Garg, Cartesia blog, "Raise Financial Services x Cartesia: Voice AI for Financial Access in India". October 12, 2026.
63. ^Cartesia customer story, "Blue Machines AI and Cartesia Partner to Bring India-Resident, Low-Latency Voice AI Solutions to Indian Enterprises".
64. ^1 ^2 ^3 ^4 ^5Cartesia, "Customers". Accessed October 2026.
65. ^Cartesia blog, "Introducing our next 8 languages on Sonic Multilingual". October 8, 2024.
66. ^Cartesia docs, "Compare TTS Endpoints". Accessed October 2026.
67. ^Cartesia docs, "Output Format". Accessed October 2026.
68. ^1 ^2Cartesia docs, "Instant Voice Clone". Accessed October 2026.
69. ^1 ^2Voximplant blog, "Extend Cartesia Line Agents to SIP, WhatsApp, and Global Phone Networks". February 11, 2026.
70. ^Albert Gu and Brandon Wang, Cartesia blog, "Hierarchical modeling". July 11, 2025.
71. ^Cartesia, "Terms of Service". Accessed October 2026.

Show all 71 references