Hume EVI

Executive Summary

Hume's Empathic Voice Interface, EVI, is a hosted API for expressive realtime conversations. Its public product page offers voice selection/design, external language models, context injection, tools and conversation-history features.^[1] It is a current option in the agentic voice APIs landscape, not merely a research demo or a text-to-speech endpoint.

Its distinctive question is how expression-aware interaction fits the application. Natural delivery can help a tutoring or conversational interface, but task correctness, interruption recovery and privacy remain separate acceptance criteria. This report reviews documentation and developer discussion, not a hands-on quality benchmark.

Versions and Architecture

These distinctions come from the current migration guide.^[2] They matter more than treating a larger version number as an unconditional upgrade. EVI 4-mini includes English, Japanese, Korean, Spanish, French, Portuguese, Italian, German, Russian, Hindi and Arabic. EVI 3 language documentation is inconsistent: the version table says English, while the FAQ also lists Spanish. Validate the intended version/language combination rather than inheriting a blanket multilingual claim.^[2][3]

Sessions use WebSocket, with messages for incoming speech/transcripts, assistant content and audio output. Authentication supports API keys or access tokens. The overview lists a thirty-minute maximum session, 16 MB WebSocket message limit and subscription-dependent concurrency.^[4] Your application supplies the surrounding user/session management and audio experience.

With a supplemental model, Hume describes sending a transcript and expression measures to that model, then voicing its text response with conversational context.^[3] This is a configurable composition; it should not be represented as one unchanging end-to-end architecture across every EVI version and configuration.

What Expression Scores Mean

The FAQ explicitly distinguishes perceived expression from inner emotional state. Scores represent the model's confidence about how vocal/language patterns might be interpreted by observers; they are not measurements of the presence or intensity of a person's actual feelings.^[3]

For a product team, that suggests a restrained use: evaluate whether a response's tone is appropriate, while checking understanding through the conversation itself. A frustrated-sounding utterance might reflect the topic, microphone conditions or speaking style. An application should not turn that score into an unsupported claim about the speaker.

Turn-Taking and Tools

Two timing controls address different events:

The first determines how long to wait after the user stops; the second determines when to yield while EVI is talking. Speech-detection sensitivity and prefix padding are additional controls.^[5][6] These are configuration values, not measured end-to-end response latencies. Lower values can react sooner while also treating pauses or brief sounds as turns.

Function tools require a compatible supplemental model. Hume's tool guide documents JSON-schema parameters, correlated tool-call IDs, responses, errors and fallback content. It also illustrates a user correcting the requested location while a lookup is in progress.^[7]

A spoken cancellation does not undo a completed external action. Separate requests that only retrieve information from requests that change a booking or account. The application should decide whether work can still be canceled, whether a replacement supersedes it, and what result the caller should hear.

A worked correction test

For a travel-information assistant, test “find the forecast for Boston,” followed immediately by “actually, Austin.” Delay the first lookup so its response arrives after the correction. Check which result is spoken, whether the outdated result remains in the transcript, and whether each tool response is associated with the right request.

Then repeat with an unavailable weather service and return an explicit failure through the documented tool-error path.^[7] The desired behavior is an honest explanation or fallback, not a confident invented forecast. This is a proposed integration test, not evidence that this review ran EVI.

Pricing and Operating Costs

The public page's EVI 3 table, checked September 16, 2026, lists these representative monthly plans:

Other tiers and enterprise terms exist. This table reports the displayed EVI 3 rates rather than assuming identical terms for every version.^[8]

The subscription also includes separate TTS/voice allowances. Hume-managed external LLM usage adds supplemental charges; with your own API key or custom model, Hume does not bill that model usage, though the provider or hosting cost still exists. Included usage resets with the monthly cycle.^[9]

As arithmetic, 1,500 EVI 3 minutes on Pro would be $70 plus 300 × $0.06, or $88 before supplemental LLM and other applicable costs. This is not a measured deployment invoice. Budget for the expected session mix and concurrency, not only an overage rate printed beside a plan.

Privacy Documentation Needs Reconciliation

Hume documents controls to disable EVI transcript/audio retention and training use; disabling retention removes history and resumption features. However, the same privacy page contains conflicting training statements: its EVI-specific section describes default use with an opt-out, while a later answer says API submissions are not used for training.^[10] The separately linked API Data Usage Policy, dated July 2023, also says customer API data is not used to train general models.^[11]

This review cannot resolve that conflict from public text. Confirm the terms and effective settings for the intended account and deployment before relying on either blanket interpretation. Configuring retention, choosing an external LLM and deciding which application logs to keep are separate parts of that review.

What Developers Say

A November 2025 React SDK issue provides a useful integration lesson. Developer digital-Chairman expected custom tool responses to appear again in the SDK's incoming-message collection. A Hume maintainer explained that built-in tool responses and application-executed functions follow different paths; the reporter tested the example and confirmed that distinction.^[12]

The discussion is evidence about that event-flow expectation, not a current universal defect or a model-quality score. A client that needs a complete business audit trail should record the function result it sends, rather than assuming every outbound application message will return through the same subscription callback.

Fit and Evaluation

EVI merits a pilot when expressive delivery, conversational pacing and configurable language-model behavior are central to the experience. Test the exact version, voice, language and supplemental model together. A pleasant voice in one short demo does not establish reliable task completion across interruptions and tool failures.

Use the voice comparison to evaluate alternative architectures on the same scenarios. Keep speech quality, answer correctness, action correctness, recovery and cost as separate measures. The public sources establish a shipped API and useful controls; they leave the noted language and privacy inconsistencies unresolved.

---

Research by Ry Walker Research • methodology