Hume AI

Hume AI is a New York-based artificial intelligence research company and API platform, founded in March 2021 by Alan Cowen, that builds "emotionally intelligent" voice AI: models trained to measure human emotional expression from speech and respond to it empathically.^[8][1] Its flagship product, the Empathic Voice Interface (EVI), is a speech-to-speech model that analyzes the tone, rhythm, and prosody of a speaker's voice to infer emotional state and adjusts its own vocal responses accordingly.^[1] Hume has released four generations of EVI (1, 2, 3, and EVI 4 mini) alongside a standalone text-to-speech model called Octave that it calls "the first LLM for text-to-speech."^[5] The company has raised $72.8 million in total funding, including a $50 million Series B led by EQT Ventures in March 2024 that Forbes reported valued it at $219 million.^[1][33] Hume operates under an ethical framework maintained by its affiliated nonprofit, the Hume Initiative on Beneficial AI.^[1][12] In January 2026 Hume agreed to license certain technologies to Google on a non-exclusive basis; Cowen joined Google, and Andrew Ettinger, previously chief revenue officer at Appen, became chief executive.^[24][25] On October 2, 2026, Hume announced that it would shut down its EVI and Octave text-to-speech APIs on November 13, 2026; the company continues, selling voice AI evaluation, expression measurement, and audio analysis products.^[28][31]

Hume occupies a specific position in the Voice AI market: it treats emotional signal not as a surface feature but as the primary training objective. Where most conversational AI systems try to sound pleasant, Hume's models attempt to detect the user's current emotional state and adjust prosody, word choice, and response tone accordingly. The company argues this produces outcomes that correlate more closely with user well-being than systems optimized purely for task completion or engagement metrics. Alan Cowen, then CEO and chief scientist, framed the approach this way: "By building AI that learns directly from proxies of human happiness, we're effectively teaching it to reconstruct human preferences from first principles and then update that knowledge with every new person it talks to and every new application it's embedded in."^[15]

History

Who is Alan Cowen and what is semantic space theory?

Alan Cowen holds a PhD in Psychology from the University of California, Berkeley, where he developed methods for mapping emotional experience using large-scale data collection and dimensionality reduction.^[8] His most widely cited paper, co-authored with Dacher Keltner and published in the Proceedings of the National Academy of Sciences in 2017, identified 27 distinct categories of reported emotional experience by having 853 participants watch 2,185 emotionally evocative short videos.^[10] That research challenged the prevailing view in psychology that emotions collapse into a handful of basic categories like happiness, sadness, and fear. Cowen and Keltner argued that human emotional experience is high-dimensional, better described by a continuous space of blended states than by discrete bins. Keltner, the study's senior author, said they found "27 distinct dimensions, not six," and "smooth gradients of emotion between, say, awe and peacefulness, horror and sadness, and amusement and adoration."^[10][23]

This work led to what Cowen calls semantic space theory, a computational framework that treats emotions as positions in a multidimensional space rather than as labeled buckets.^[11] The theory holds that the nuances of voice, face, body movement, and gesture all carry meaning that standard categorical emotion labels miss.^[11] Cowen's research has appeared in Nature Human Behaviour and Science Advances, and Contrary Research reported that Hume had published more than 40 research articles with over 3,000 citations as of January 2026.^[8] Semantic space theory remains the scientific foundation of Hume's affective computing models.

After completing his PhD, Cowen joined Google AI, where he led the Affective Computing team.^[8] At Google he continued studying how AI systems could be trained to perceive emotional signals across modalities. According to Contrary Research, he grew increasingly concerned that AI technologies such as voice assistants and large language models lacked sensitivity to human emotion and well-being.^[8] This tension between what emotion-sensing AI could do to help people versus what it could do to manipulate them became the central framing for Hume AI's founding mission.

When was Hume AI founded?

Cowen left Google in 2021 and founded Hume AI in March of that year.^[8] The company's name references the Scottish philosopher David Hume, whose writing on emotions as the primary drivers of human choice and welfare informed Cowen's framing. From the start, Hume positioned itself as a research lab as much as a product company, publishing scientific work alongside building commercial tools.

The founding team included John Beadle, co-founder and managing partner at Aegis Ventures, who joined as a founding investor, CFO, and board member, and Janet Ho as COO.^[8] Ho had previously served as managing partner at Aegis Ventures and brought operational experience from Zynga and Rakuten.^[8]

In the early period Hume released its Expression Measurement API, which allowed developers to submit audio recordings, video files, images, and text to receive structured output describing the emotional dimensions detected in each modality.^[8] In September 2022 the company began rolling out a beta version of that technology to a waitlist of more than 2,000 companies and research organizations.^[7] The early customer mix reflected Hume's research-lab origins: academic medical centers interested in clinical monitoring, psychology researchers studying human expression at scale, and a smaller set of commercial developers building health and wellness applications.

On January 22, 2026, Hume announced that Andrew Ettinger had been appointed chief executive. The same release said Hume had agreed to non-exclusively license certain technologies to Google and that co-founder Alan Cowen had joined Google. Ettinger, who had been chief revenue officer at Appen, said that separately from the Google agreement Hume was on track to generate more than $100 million in revenue that year from research partnerships with frontier AI labs and AI-first enterprises.^[24] Wired reported that Google DeepMind was also hiring several of Hume's top engineers to help bring voice and emotion technology into its frontier models.^[25] Cowen was a co-author of Google's September 23, 2026 launch post for Gemini 3.8 Flash TTS, signing as a director of research science on behalf of the Gemini Audio team.^[26] The next day Hume published results from its Real World VoiceEQ benchmark, which tested preview versions of the Google models: it ranked Gemini 3.8 Flash TTS first overall in Voice Design and first on its Expressivity-Reliability Frontier Score, but eleventh of 13 models on speaker similarity for voice replication. The post disclosed that Hume has a non-exclusive licensing agreement with Google and said its held-out test set is not shared with model developers.^[27]

How much funding has Hume AI raised?

Hume raised a $12.7 million Series A round in January 2023, led by Union Square Ventures.^[7] Additional participants included Comcast Ventures, LG Technology Ventures, Northwell Holdings, Wisdom Ventures, and Evan Sharp, co-founder of Pinterest.^[7] Coverage quoted in Hume's announcement said the company intended to use the funds to meet demand for its technology, which analyzed human expressive behavior in images, audio, video, and text.^[7] At the time, Hume had research partnerships with labs at Mount Sinai, Boston University Medical Center, and Harvard Medical School examining how analysis of vocal and facial expression could improve healthcare outcomes.^[7]

In March 2024, Hume announced a $50 million Series B round led by EQT Ventures.^[1] Other participants included Union Square Ventures (returning), Nat Friedman and Daniel Gross, Metaplanet, Northwell Holdings, Comcast Ventures, and LG Technology Ventures.^[1] Forbes reported that the round valued the company at $219 million.^[33] The announcement came on March 25, 2024, simultaneously with the public unveiling of EVI.^[1][15] Total funding reached $72.8 million.^[1][8]

The table below summarizes Hume's disclosed funding rounds.

The participation of Northwell Holdings, the investment arm of Northwell Health (one of the largest US hospital systems), was notable because it signaled healthcare as a target vertical.^[1] Northwell's involvement gave Hume a potential channel for clinical deployments and framed the investment as strategic rather than purely financial.

By early 2025 Hume reported more than 200 new API platform sign-ups per week, and as of November 2025 it said it had more than 100,000 customers across developers and businesses; Contrary Research's October 2026 profile listed 56 employees.^[8]

Empathic Voice Interface

The Empathic Voice Interface was Hume's core commercial product from 2024 until Hume announced in October 2026 that the EVI API would shut down on November 13, 2026.^[28] It is a speech-to-speech AI that handles the full pipeline of a voice conversation: transcribing incoming audio, understanding language and emotional context, generating a response, and synthesizing speech with prosody tuned to the situation.^[1] Hume describes EVI as "a first-of-its-kind conversational AI with emotional intelligence" that is "trained on data from millions of human interactions and uses vocal tones to understand when users finish speaking, predict their preferences, and optimize responses for satisfaction over time."^[1] Each generation of EVI has been released alongside API access so developers can embed it in their own applications.

EVI 1 (2024)

Hume unveiled EVI on March 25, 2024, alongside the Series B announcement, and said it would be publicly available in April 2024.^[1] EVI 1 introduced several capabilities that distinguished it from conventional voice assistants at the time.

The system used what Hume called an empathic large language model (eLLM), a multimodal architecture that combined a large language model with the company's expression measurement technology.^[1] Hume describes the eLLM as "a new form of multimodal generative AI that integrates large language models (LLMs) with expression measures," enabling EVI to "adjust the words it uses and its tone of voice based on the context and the user's emotional expressions."^[1] Rather than treating speech purely as text input, the eLLM received the audio directly and used Hume's trained models to extract emotional signal from prosody before generating a response.^[1]

EVI 1 also introduced intelligent end-of-turn detection. Conventional voice systems wait for silence before responding, which leads to awkward pauses when speakers trail off without finishing a thought. EVI's approach used vocal tone as an additional signal to infer when a speaker had genuinely finished rather than pausing mid-sentence.^[1]

The system could be interrupted mid-response, stopping speech when the user spoke over it.^[1] It generated responses in a voice that could vary in tone, pace, and pitch in response to the user's detected emotional state.^[1] Hume described this as the model learning to optimize for user satisfaction by adjusting its own expressive behavior based on how users responded.^[1]

At launch, EVI 1 was available through a WebSocket API. Developers could configure the underlying language model (choosing from providers including Anthropic, OpenAI, and others), set a system prompt to define the persona and task, and connect custom tools.^[1] Hume also released SDKs for Python and TypeScript to simplify integration.^[1]

The April 2024 public availability of EVI 1 put Hume into direct competition with OpenAI's then-unreleased GPT-4o voice mode, which OpenAI demonstrated publicly in May 2024. The two products addressed overlapping problems but with different architectures and different emphases. EVI 1 foregrounded emotional responsiveness and LLM flexibility; GPT-4o voice prioritized natural-sounding speech and tight integration with the OpenAI model ecosystem.

EVI 2 (September 2024)

Hume released EVI 2 in September 2024 in beta.^[2] Hume described it as a new voice-to-voice foundation model.^[2]

EVI 2 introduced continuous voice modulation without voice cloning. Developers could adjust parameters along scales including pitch, gender presentation, and nasality to create custom voice characteristics.^[2] The architecture deliberately prevented voice cloning: by design, EVI 2 could not replicate a specific person's voice without modifications to its underlying code.^[2] Hume cited safety considerations in this decision, aiming to prevent the system from being used to impersonate real individuals.^[2]

The model achieved sub-second response latency and supported what Hume called emergent multilingual capabilities, with initial fluency in English, Spanish, French, German, and Polish.^[14] The model released that day was EVI-2-small; Hume said it was fine-tuning a larger EVI-2-large to announce later.^[2]

In a comparison Hume published against OpenAI's GPT-4o voice model, EVI 2 was noted to cost approximately $4.32 per hour versus $9 per hour for GPT-4o Realtime, while offering flexibility to use any LLM provider rather than being locked to OpenAI's models.^[14] EVI 2 also ran with more personality customization options: the continuous voice modulation parameters let developers tune voice characteristics without needing a new voice clone for each variation, which made it practical to build products offering multiple distinct AI personas without a large upfront data collection effort.^[14]

EVI 3 (2025)

Hume announced EVI 3 on May 29, 2025, initially through a live demo and iOS app, with API access made available in the weeks that followed.^[3][4] EVI 3 represented a more substantial architectural shift. Where earlier versions of EVI separated language understanding and speech synthesis into distinct components, EVI 3 was designed as a unified speech-language model in which "a single autoregressive model processes both text (T) and voice (V) tokens" in one pass.^[3]

With the API release in July 2025, Hume added voice cloning: using 30 seconds of audio or less, EVI 3 could capture not only a voice's timbre and accent but also its rhythm, tone, and "even aspects of the speaker's personality."^[4] At launch the system had access to more than 100,000 custom voices built on the Octave TTS platform, each with an inferred personality profile; later API announcements raised that figure to over 200,000.^[3][4]

Voice and personality could be specified entirely through natural language prompts, without fine-tuning.^[3] A developer could write a prompt describing a character's voice, speech patterns, and emotional tendencies, and EVI 3 would generate speech matching that description.^[3]

EVI 3 supported integration with external language models including Claude 4 (Anthropic), Gemini 2.5 (Google), and Kimi K2.^[4] Developers could also connect custom LLMs or retrieval-augmented generation pipelines.^[4] The architecture generated initial responses while external LLMs processed in parallel, then handed off seamlessly to the external model's output once available.^[4]

In blind comparisons Hume conducted with human raters, EVI 3 was rated higher than GPT-4o across all seven dimensions evaluated: amusement, audio quality, empathy, expressiveness, interruption handling, naturalness, and response speed.^[3] It also outperformed Gemini and Sesame (AI company) on emotion and style modulation across 30 distinct speaking styles, and identified eight of nine tested emotions more accurately than GPT-4o.^[3] On practical latency measured from New York, Hume reported EVI 3 averaging about 1.2 seconds, versus about 2.6 seconds for GPT-4o, 1.5 seconds for Gemini, and 1 second for Sesame, and said EVI 3 could respond in under 300 milliseconds on state-of-the-art hardware.^[3]

When the EVI 3 API became available on July 18, 2025, Hume announced that EVI 1 and EVI 2 would be deprecated on August 30, 2025, giving developers about six weeks to migrate.^[28]

EVI 3 also integrated with the Hume app, a consumer iOS application that gave end users direct access to a conversational AI companion without developer intermediation.^[3] The app was positioned as a personal AI product rather than a developer tool, targeting general consumers interested in voice-based AI interaction.

EVI 4 mini (October 2025)

In October 2025, alongside the release of Octave 2, Hume launched EVI 4 mini.^[9] The model uses Octave 2 as its speech synthesis layer and requires pairing with an external large language model for full language generation.^[9] It targets use cases where low latency and low cost matter more than the full integrated personality and emotional reasoning of EVI 3. EVI 4 mini is available through the Hume API with WebSocket streaming support.^[9]

What is an empathic large language model (eLLM)?

The empathic large language model (eLLM) is Hume's term for the multimodal architecture that powers EVI. The concept was introduced publicly with the launch of EVI 1 in 2024.^[1]

Conventional speech AI pipelines chain together separate systems: a speech-to-text transcriber, a language model, and a text-to-speech synthesizer. Each component receives only the output of the previous one. By the time language reaches the LLM, paralinguistic information about how words were spoken has been discarded.

Hume's eLLM architecture instead feeds the audio signal directly into a system that jointly processes language content and expression measures derived from prosodic analysis.^[1] This lets the model condition its response on detected emotional state. If a speaker sounds frustrated, the model can acknowledge that frustration in its word choice and in the prosody of its synthesized response. If the user's tone brightens, the model can reflect that change.

In practice, the eLLM uses output from Hume's expression measurement technology as additional context for language generation. The model is trained using feedback derived from how users respond to different types of AI output, treating proxies of user satisfaction as the optimization target rather than relying solely on human preference ratings collected in controlled studies.^[1]

With EVI 3, the eLLM concept evolved further into a fully unified speech-language model where speech encoding, language modeling, and speech decoding are handled in a single architecture rather than as separate modules sharing outputs.^[3]

One consequential implication of the eLLM design is that the system can generate initial words while simultaneously processing the rest of a response from an external LLM.^[4] This speculative generation reduces perceived latency: the user hears a response beginning almost immediately while the full answer is still being computed. The architecture is somewhat analogous to speculative decoding in text generation, adapted for the sequential constraints of audio output.

Octave

Octave is Hume's text-to-speech product, released separately from EVI to serve use cases that do not require two-way conversation. Hume describes Octave as "the first LLM for text-to-speech," meaning that the system is built on a language model architecture rather than the signal-processing pipelines used by conventional TTS systems.^[5]

Octave 1 (February 2025)

Hume released Octave on February 26, 2025.^[5] The core claim was that Octave understood the meaning and emotional context of the text it was reading, rather than simply converting characters to sound: Hume said Octave "interprets plot twists, emotional cues, and character traits within a script or prompt, then transforms that understanding into lifelike speech."^[5] A conventional TTS system reads words in sequence without knowledge of narrative context or character intent. Octave's LLM foundation allowed it to infer how lines should be delivered based on what the words mean and how they fit the surrounding passage.^[5]

Octave allowed developers to specify voices through natural language descriptions, ranging from simple directives like an English accent to detailed character briefs like a goblin auctioneer with a deep, gruff, gravelly voice.^[5] Developers could also give acting instructions to control emotional delivery: telling the system to deliver a line in a whispered and hushed tone, or angry and furious.^[5][16] Speed and pause parameters were adjustable as well.

At launch, Octave offered more than 60 pre-made voices, audio output at 48kHz, and a Creator Studio interface for producing long-form content such as audiobooks and podcasts.^[5]

In benchmark tests against ElevenLabs Voice Design using 180 human raters evaluating 120 diverse prompts, Octave was preferred on audio quality 71.6% of the time, on naturalness 51.7% of the time, and on prompt adherence 57.7% of the time.^[5]

Octave 2 (October 2025)

Octave 2 launched on October 1, 2025, alongside EVI 4 mini.^[6][9] The update delivered a 40% reduction in latency (responses under 200 milliseconds) and a 50% reduction in cost compared to Octave 1, achieved partly through a compute optimization partnership with SambaNova.^[6] Octave 2 added support for 11 languages: Arabic, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.^[6] It also introduced multi-speaker conversation capability, enhanced reliability on uncommon words and numbers, and a voice conversion feature that could replace one speaker's voice with another while preserving the original phonetic timing, useful for dubbing applications.^[6]

Expression Measurement API

The Expression Measurement API was Hume's earliest commercial product and the foundation for its early research partnerships.^[8] The API accepted audio, video, images, and text and returned structured measurements of emotional expression across multiple modalities.^[8]

The system measured 48 dimensions of emotional expression from facial movements, 48 dimensions from vocal prosody (tone, rhythm, and timbre of speech), 48 dimensions from non-linguistic vocalizations such as laughs, sighs, and gasps, and 53 dimensions of emotional expression from the meaning and tone of text. These dimensions were derived from Cowen's semantic space theory research and the 27-emotion taxonomy from his Berkeley work, extended and refined through further studies.^[11][10]

The API supported batch processing through a job-based interface as well as streaming. Use cases included research applications at academic medical centers, mental health assessment tools, and customer experience analytics.

Hume sunset the original Expression Measurement API in mid-2026: the last day to create new jobs through the API Playground was May 14, 2026, API access ended on June 14, 2026, and Hume's SDKs removed it after the sunset.^[30] On October 6, 2026, Hume launched a new Expression Measurement API in limited release, with audio and video endpoints for uploaded media and live streams.^[29]

The sunset was a significant change for researchers and clinical developers who had built workflows around the original batch-processing API.

API and pricing

Hume distributes all its products through a unified API platform.^[18] Access requires an API key and uses WebSocket connections for real-time voice interactions (EVI) and standard HTTP for Octave TTS.

Pricing uses a tiered subscription model with monthly allotments of usage and overage rates for usage beyond the included amounts.^[18]

For EVI (speech-to-speech), Hume's pricing page listed the following plans from late 2025 into 2026 (it no longer publishes prices):^[18]

For Octave TTS (text-to-speech), pricing was based on character volume:

Expression Measurement (while active) used pay-as-you-go rates: $0.0828 per minute for video with audio, $0.0639 per minute for audio only, $0.045 per minute for video only, $0.00204 per image, and $0.00024 per word for text.^[18]

Hume also offered startup grants and volume discounts for large-scale deployments.^[18] At high enough volume, Hume said the per-minute rate on EVI could fall well below $0.02.^[4]

What is Hume AI used for?

Mental health and wellness

Several of Hume's documented customers operate in mental health and wellness. Ream, a mental health service for adults with ADHD, integrated EVI to enable daily five-minute AI coaching calls.^[19] The service found that voice-based coaching sessions powered by EVI doubled daily active users compared to text-based sessions.^[19] The company noted that EVI's emotional responsiveness contributed to user retention, with users expressing a sense of genuine connection to the coaching interaction.^[19]

hpy, a therapy platform, used both the Expression Measurement API and EVI in combination.^[20] Therapists received structured output from expression measurement during sessions, giving them data on patterns in clients' vocal and facial expression over time.^[20] Between sessions, clients could use EVI as a conversational companion providing therapy-informed support.^[20]

Coaching and performance

The Hume Initiative's list of supported use cases includes athletic and professional performance optimization.^[13] Developers have built coaching applications using EVI that deliver feedback calibrated to the learner's emotional state, adjusting tone when frustration or disengagement is detected rather than continuing to deliver information in a flat instructional register.

Customer support and service

EVI's end-of-turn detection and interrupt handling make it useful for automated phone support, where natural conversation flow is important for user experience. Hume has positioned EVI as a tool for building voice agents in customer service contexts, with the emotional responsiveness providing a warmer interaction quality than systems that do not attend to vocal tone.

Developer tools and character animation

The Hume API is also used by developers building social AI companions, interactive fiction, and character animation. EVI 3's voice and personality cloning capabilities, combined with the ability to specify character voice through natural language prompts, have made it a tool for interactive entertainment applications where a large cast of distinct-sounding characters is needed without manual fine-tuning.

Healthcare and research

Hume's early research partnerships with Mount Sinai, Boston University Medical Center, and Harvard Medical School focused on applying expression measurement to clinical contexts: standardized patient screening, mental health diagnosis support, and monitoring patient emotional state during treatment.^[7] These applications used the Expression Measurement API rather than EVI.

Hume Initiative on Beneficial AI

The Hume Initiative is a nonprofit organization founded alongside Hume AI to develop ethical guidelines for empathic AI systems.^[12] The Initiative's website lists Cowen on the ethics committee that drafted its guidelines.^[12] The initiative operates independently of the commercial company, though Cowen, who founded both, gave up the company's chief executive role when he joined Google in January 2026.^[24]

The initiative convened an ethics committee of 11 experts spanning AI research, emotion science, legal practice, privacy law, public policy, and industry standards.^[12] Members include Ben Bland, chair of the IEEE P7014 Empathic AI Working Group; Danielle Krettek-Cobb, founder of Google Empathy Lab; Kristen Mathews from Morrison and Foerster; and Edward Dove from the University of Edinburgh.^[12]

The initiative's central concern is that AI systems capable of detecting and responding to emotional state can be used in ways that exploit vulnerability rather than support well-being.^[12] Current AI systems often optimize for engagement signals that do not correlate with user welfare, and emotion-aware AI could deepen this problem. The guidelines aim to distinguish between applications that improve human emotional experience and those that manipulate or exploit it.^[13]

Supported use cases in the initiative's framework include mental health applications, empathic digital assistants optimized for well-being, accessibility tools such as speech emotion transcription, health monitoring and crisis assessment, and social algorithms designed to surface content that is genuinely satisfying rather than merely habit-forming.^[13] The initiative's unsupported use cases are manipulation, deception, optimizing for reduced well-being, and "unbounded empathic AI"; its About page warns that empathic algorithms could "surface and reinforce unhealthy temptations when we are most vulnerable to them."^[12]

Hume has said that all of its work is guided by the initiative's ethical guidelines, and that it has committed to specific use cases it will never support.^[7]

The initiative's work addresses a regulatory area that law has only begun to cover. The European Union's AI Act bans AI systems that infer emotions in workplaces and educational institutions (except for medical or safety reasons), a prohibition in force since February 2, 2025, and classifies most other emotion recognition systems as high-risk.^[32] The Hume Initiative's framework predates these regulatory developments and was designed to provide enforceable guidelines while waiting for law to catch up.

How does Hume AI compare with other voice AI platforms?

The following table compares Hume AI's EVI with other major real-time voice AI platforms as of mid-2025.

In comparisons conducted by Hume using blind human ratings, EVI 3 was rated higher than GPT-4o on all seven dimensions tested, and outperformed Sesame (AI company) on emotion and style modulation across 30 distinct styles.^[3]

Limitations

EVI's emotional inference is probabilistic. The model estimates emotional state from acoustic signals and does not have direct access to what a user is actually feeling. Prosodic cues can be ambiguous: a flat tone might indicate boredom, fatigue, or simply a reserved speaking style. The system can misread emotional context, particularly across cultural variation in prosodic norms.

Despite the Hume Initiative's ethical framework, the same technology that enables supportive emotional responsiveness could be adapted to manipulative purposes by actors who deploy it outside the initiative's guidelines. Hume's use policies create contractual constraints but cannot prevent misuse by parties who agree to those policies and then violate them.

The original Expression Measurement API, which provided the most detailed structured output of emotional signal across modalities, was sunset in June 2026; its October 2026 replacement launched only in limited release, measuring speech and faces rather than text.^[30][29]

EVI 3's voice and personality cloning capabilities, while subject to Hume's policies, raise questions about consent and impersonation that the company acknowledges but has not fully resolved. The 30-second cloning threshold is low enough that a voice sample obtained without a subject's awareness could in principle be used to generate cloned output. Cowen has publicly warned about the risks of celebrity impersonation and the need for safeguards as AI voice cloning becomes widely available.^[21]

As of mid-2025, EVI 3 was primarily tested and optimized for English, with French, German, Italian, and Spanish planned, while OpenAI's Realtime API and ElevenLabs covered a wider range of languages.^[3] Octave 2 and EVI 4 mini extended Hume's voice models to 11 languages in October 2025.^[6]

Continue reading

• OpenAIOpenAI is an American artificial intelligence research and deployment organization.Mentioned 11 times above
• TText-to-SpeechText-to-speech (TTS) refers to artificial intelligence systems that convert written text into natural-sounding spoken audio.Mentioned 7 times above
• VVoice AIVoice AI is the umbrella term for artificial intelligence systems that listen to, understand, and generate human speech.Mentioned 5 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 3 times above
• Sesame (AI company)Sesame (formally Sesame AI Labs) is a San Francisco-based artificial intelligence company founded in June 2023, best known for developing the Conversational Speech Model (CSM) and the Maya and Miles voice…Mentioned 3 times above
• CartesiaCartesia is a San Francisco-based AI company focused on real-time voice synthesis, speech recognition, and state space model (SSM) research.Listed in See also

More from See also

• Affective computing
• Conversational Models

More to read

• Google
• Artificial Intelligence
• Voice cloning
• Gemini 3.8 Flash TTS

Topics:AI CompaniesConversational AISpeech & Audio AIVoice AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24Hume AI. "Hume Raises $50M Series B and Releases New Empathic Voice Interface." hume.ai/blog/series-b-evi-announcement. March 25, 2024.
2. ^1 ^2 ^3 ^4 ^5 ^6Hume AI. "Introducing EVI 2, our new foundational AI voice model." hume.ai/blog/introducing-evi2. September 2024.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Hume AI. "Introducing EVI 3: the world's most realistic and instructible speech-to-speech foundation model." hume.ai/blog/introducing-evi-3. May 29, 2025.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Hume AI. "Announcing EVI 3 API: The most customizable speech-to-speech model." hume.ai/blog/announcing-evi-3-api. July 17, 2025.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Hume AI. "Octave TTS: the first text-to-speech system that understands what it's saying." hume.ai/blog/octave-the-first-text-to-speech-model-that-understands-what-its-saying. February 26, 2025.
6. ^1 ^2 ^3 ^4 ^5Hume AI. "Octave 2: next-generation multilingual voice AI." hume.ai/blog/octave-2-launch. October 2025.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7Hume AI. "Hume AI Raises $12.7M in Series A Funding." hume.ai/blog/hume-ai-raises-usd12-7m-in-series-a-funding. February 16, 2023.
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Contrary Research. "Hume AI Business Breakdown & Founding Story." research.contrary.com/company/hume-ai.
9. ^1 ^2 ^3 ^4Testing Catalog. "Hume AI launches Octave 2 and EVI 4 mini voice models." testingcatalog.com. October 2025.
10. ^1 ^2 ^3Cowen, Alan S., and Dacher Keltner. "Self-report captures 27 distinct categories of emotion bridged by continuous gradients." *Proceedings of the National Academy of Sciences* 114, no. 38 (2017): E7900-E7909.
11. ^1 ^2 ^3Cowen, Alan, et al. "Semantic Space Theory: A Computational Approach to Emotion." *Trends in Cognitive Sciences* 25, no. 2 (2021): 124-136.
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7Hume Initiative. "About." thehumeinitiative.org/about/.
13. ^1 ^2 ^3Hume Initiative. "Supported Use Cases." thehumeinitiative.org/supported-use-cases/.
14. ^1 ^2 ^3Hume AI. "Comparing the world's first voice-to-voice AI models: EVI 2 and GPT-4o." hume.ai/blog/evi2-vs-gpt4ovoice.
15. ^1 ^2BusinessWire. "Hume AI Announces $50 Million Fundraise and Empathic Voice Interface." businesswire.com. March 26, 2024.
16. ^VentureBeat. "Hume launches new text-to-speech model Octave that generates custom AI voices with adjustable emotions." venturebeat.com. February 2025.
17. SurePrompts. "Voice Generation Models Compared (2026): ElevenLabs, OpenAI TTS, Hume, Cartesia, PlayHT." sureprompts.com.
18. ^1 ^2 ^3 ^4 ^5Hume AI. "Pricing." hume.ai/pricing. Archived December 2025: web.archive.org/...pricing
19. ^1 ^2 ^3Hume AI. "Voice AI for Mental Health Coaching." hume.ai/blog/case-study-hume-ream-app-limited.
20. ^1 ^2 ^3Hume AI. "How hpy Uses Hume's AI voice to Enhance Mental Health Care." hume.ai/blog/case-study-hume-hpy.
21. ^Tom's Guide. "AI voice cloning is exploding: Hume's CEO warns about celebrity impersonations and the urgent need for safeguards." tomsguide.com. 2025.
22. Maginative. "Hume AI Raises $50M Series B, Unveils Empathic Voice Interface." maginative.com. March 27, 2024.
23. ^Yasmin Anwar. "Researchers pinpoint 27 states of emotion." Berkeley News. September 6, 2017. news.berkeley.edu/...27-emotions
24. ^1 ^2 ^3Hume AI via PR Newswire. "Hume AI Appoints New CEO." January 22, 2026. prnewswire.com/...hume-ai-appoints-new-ceo-302668103
25. ^1 ^2PYMNTS. "Google Recruits Hume CEO Alan Cowen to Bolster Voice AI Efforts" (reporting on Wired). January 22, 2026. pymnts.com/...-alan-cowen-bolster-voice-ai-efforts
26. ^Leland Rechis and Alan Cowen. "Gemini 3.8 text-to-speech says hello." Google blog. September 23, 2026. blog.google/...gemini-3-8-text-to-speech
27. ^Alice Baird. "Newly-released Google's Gemini 3.8 Flash TTS tops Hume's Real-World VoiceEQ leaderboard." Hume AI blog. September 24, 2026. hume.ai/...s-hume-s-real-world-voiceeq-leaderboard
28. ^1 ^2 ^3Hume AI. "Changelog" (Voice APIs), including "TTS and EVI API sunset notice" (October 2, 2026) and "EVI API changes" (July 18, 2025). Hume developer documentation. Archived: web.archive.org/...changelog
29. ^1 ^2Hume AI. "Expression Measurement API launch." Expression Measurement changelog. October 6, 2026. dev.hume.ai/...6.md
30. ^1 ^2HumeAI. "hume-python-sdk v0.14.0" release notes. GitHub. 2026. github.com/...v0.14.0
31. ^Hume AI. Homepage and product list. Accessed November 2026. hume.ai
32. ^European Union. "Regulation (EU) 2024/1689 (Artificial Intelligence Act)." Official Journal of the European Union. July 12, 2024. eur-lex.europa.eu/...oj
33. ^1 ^2 ^3Rashi Shrivastava. Forbes report on Hume's $50M Series B at a $219M valuation, March 2024, as aggregated: news.bensbites.co/...out

Show all 33 references