Hume vs ElevenLabs now hinges on what you need to keep running. Hume's announced voice API sunset changes the decision for both new buyers and existing integrations, while its newer audio-analysis services serve a different purpose. The meaningful comparison separates speech generation, live conversation and audio evaluation, then accounts for model controls, voice assets, streaming protocols and the costs of replacing each layer.
The buying decision has changed
Hume's announced cutoff for Octave text-to-speech and Empathic Voice Interface access has passed. Actual endpoint and account behavior remains unconfirmed. ElevenLabs offers speech generation and voice-agent products. Hume's newer analysis APIs do a different job.
Hume's October 2, 2026 API sunset notice set the end of text-to-speech (TTS) and Empathic Voice Interface (EVI) access for November 13, 2026, at 05:01 UTC. It also warned that account data would be permanently deleted after November 13. The notice did not identify a deletion time or enumerate the affected data.
That was an announced timetable. It does not establish the current state of individual voice endpoints, accounts or exports, or whether data deletion has occurred. The old Voice documentation now redirects to Audio Decisions. Those redirects confirm neither shutdown nor continued API access. Historical support promises and feature descriptions cannot guarantee present access.
For a new speech dependency, Hume's announced end of TTS/EVI access rules it out as an ongoing foundation. For an existing user, the relevant choice is the replacement scope:
| Your requirement | Hume's position | ElevenLabs' corresponding scope |
|---|---|---|
| Generate speech from text | Octave TTS's announced access cutoff has passed | TTS models, including v4 and v4 Turbo |
| Run a spoken conversation | EVI's announced access cutoff has also passed | ElevenAgents, with Custom mode or Realtime models in a gradual rollout |
| Judge recorded speech or conversational behavior | Audio Decisions and Expression Measurement serve analysis needs | Separate from buying a synthesis model or agent runtime |
An Octave integration depends on speech output. An EVI integration also depends on a conversational interface around speech input and responses. Replacing the first changes an output component; replacing the second also changes conversation behavior.
Hume's current services analyze audio rather than replace Octave
The TTS/EVI notice does not announce a company-wide shutdown. Its current developer catalog centers on Audio Decisions and Expression Measurement.
The Audio Decisions API takes an audio file and questions, then returns structured probabilities. A question can concern naturalness, a caller's tone, background sounds or an event within a recording. It returns no generated speech, transcript or conversational reply.
Its current file interface accepts recordings from one second to five minutes. New users are directed to a product access waitlist, so a published API contract does not imply unrestricted self-service access. Live and ambient uses remain a developing opportunity, distinct from the available file-analysis interface.
Expression Measurement scores vocal and facial expression through upload and real-time interfaces. These are measurements of expressed or perceived signals, not direct knowledge of a person's internal feelings. The newer services use the Hume Platform. Under its current authentication rules, old app.hume.ai keys do not work with the current Expression Measurement API.
This creates a complementary architecture: ElevenLabs can generate a voice, while Hume can analyze a recording of it. Neither analysis service supplies the speaking voice or conversation runtime that an Octave or EVI application needs to replace.
Speech models, languages and expression controls
For synthesis, the useful comparison is now between the requirements of an existing Hume integration and a specific replacement model. A company-wide language count or a historical preview feature list is too broad.
ElevenLabs' speech model family has different language and generation boundaries:
| Model | Languages | Relevant use |
|---|---|---|
| Eleven v4 | 90+ | Expressive content and dialogue generation |
| Eleven v4 Turbo | 90+ | Interactive speech generation |
| Eleven v3 / v3 Conversational | 70+ | Earlier expressive model family |
| Eleven Multilingual v2 | 29 | Long-form speech generation |
| Eleven Flash v2.5 | 32 | Low-latency speech generation |
V4 and v4 Turbo have model-specific generation controls: audio tags, Stability and Similarity settings. Professional Voice Clones using Eleven v4 also have a Clone Strength control. Style and Speed sliders are unavailable in v4, and it does not support Speech Synthesis Markup Language (SSML).
An Octave acting description therefore becomes a performance brief rather than a parameter that transfers unchanged. Delivery directions need a model-specific representation. A tag such as [whispering] guides generation; it does not guarantee an identical performance on every request.
V4 also changes cross-language accent behavior. When the output language differs from the reference recording's language, it aims for native-sounding target-language speech rather than carrying the original-language accent across. A multilingual brand voice can retain recognizable characteristics while losing an accent associated with the original speaker.
For a booking assistant, intelligible dates, names and mixed-language addresses matter more than the total language count. For a character, recognizable delivery and accent can be part of the product itself. Neither case is resolved by a provider-wide quality verdict.
Voice cloning: recordings are more portable than voice IDs
A Hume voice ID identifies an asset inside Hume. It is not an ElevenLabs voice model. The reusable inputs are authorized source recordings, speaker permissions, pronunciation requirements and the intended delivery.
ElevenLabs has two cloning workflows:
- Instant Voice Cloning uses short samples. Its guidance recommends about one to two minutes of clean, consistent audio.
- Professional Voice Cloning (PVC) adds training and speaker verification. Its minimum recording recommendation is 30 minutes, with two to three hours preferred. PVC requires an eligible subscription, starting at Creator.
Recording length is guidance, not a fidelity or completion-time guarantee. Under the professional cloning rules, a speaker must create and verify their own PVC. Another person's consent alone does not let a developer create it on their behalf; the speaker can create it and share it privately.
Even an ElevenLabs clone cannot be exported as a standalone voice model. Reusing source audio across providers also does not produce identical output. A replacement project includes voice recreation and acceptance of the new delivery, alongside the API changes.
Hume's announced account-data deletion makes provider-held copies a poor sole archive. Needed recordings, generated audio and configuration records have value independently of the voice IDs that reference them.
Streaming requires a protocol change
Output streaming and incremental text input solve different problems. HTTP streaming returns audio while generation proceeds. A WebSocket can also accept text as a language model produces it. Playback, buffering, cancellation and late audio still need an owner in the application.
ElevenLabs has two speech WebSocket families:
| Interface | Model scope | Request structure |
|---|---|---|
| TTS WebSocket | Flash, Multilingual v2 and other supported TTS models; excludes v3/v4 | Voice in the URL, followed by text messages |
| Text to Dialogue WebSocket | Supported v3/v4-family models | Initial voices registration, followed by inputs with a voice_id |
Text to Dialogue connections occupy a separate dialogue-session pool throughout their open lifetime. An idle socket can consume capacity. Standard TTS WebSocket concurrency counts active generation instead.
Transport limits also differ from headline model limits. The HTTP dialogue endpoint advises no more than 2,000 combined text characters per request for reliable generation. That boundary does not become larger because a model card advertises a higher character limit.
An Octave adapter replacement therefore includes message formats, audio encoding, sample rate, end-of-generation signals and interruption cleanup. An EVI replacement includes those changes plus speech input, turn boundaries, response generation, tool events and conversation state.
ElevenLabs has agent options beyond TTS
Treating ElevenLabs as only a speech-output API would miss a major part of the current comparison.
ElevenAgents supplies a hosted agent platform. In Custom mode, teams select component settings, including the language model and TTS model. Eleven Realtime v1 Mini and v1 Max offer a different operating model: ElevenLabs selects and manages the speech-recognition, language-model and synthesis components together.
The Realtime rollout limits matter. Access is gradual across workspaces. Realtime mode does not permit a custom language model or workflow overrides of managed component settings. It also restricts transfers between Realtime and Custom agents. A requirement to retain a particular model can therefore favor Custom mode.
The direct Realtime API has an OpenAI-compatible WebSocket interface. Its documented browser flow selects behavior through a model value and initial session configuration, rather than a dashboard agent_id. That protocol is not Hume's EVI protocol. Realtime uses a cascaded architecture with distinct models, rather than a single fused speech model.
For an EVI user, there are consequently two decisions: which voice to recreate and which system will own the conversation. Moving to standalone TTS leaves the second decision unresolved. Moving to an integrated agent configuration ties more of the stack to one provider and makes a later migration broader than a synthesis-adapter change.
Latency figures do not establish a winner
ElevenLabs reports approximately 100 milliseconds median server-side first-audio latency for v4 Turbo and Flash v2.5. Its latency measurement definition, dated October 29, 2026, uses production WebSocket traffic and warm voices. It excludes network travel, upstream application work and playback buffering.
That synthesis number is neither a complete-agent response time nor a guarantee for every request. Historical Octave preview figures cannot establish current Hume performance or a like-for-like tie.
A caller experiences the whole path: the end of their turn, response generation, synthesis, transport and playback. Replacing EVI with separately assembled components changes that path even when the chosen TTS model has a low first-audio number.
Costs depend on which layer you are buying
Hume's announced end of voice API access makes a historical voice subscription unsuitable as a continuing price benchmark against ElevenLabs synthesis. The relevant costs are the replacement service, voice recreation and application changes.
ElevenLabs' base API synthesis rates are:
| Model group | Per 1,000 characters | One million characters at that base rate |
|---|---|---|
| v4, v3, Multilingual v2 | $0.08 | $80 |
| v4 Turbo, v3 Conversational, Flash / Turbo | $0.04 | $40 |
These are character-based synthesis calculations before taxes. Spaces, punctuation and audio tags count toward usage. Some shared voices carry credit multipliers, so the base-rate calculation is not a universal invoice prediction.
Conversation products have separate pricing. ElevenAgents' call-minute rates include additional minutes at $0.08 and burst minutes at $0.16, plus language-model usage. External carrier charges remain separate.
The direct Realtime API lists $0.08 per additional connected audio minute, plus model-token charges. Mini lists $0.20 per million input tokens and $6 per million output tokens; Max lists $1.80 and $9, respectively. A TTS-character budget cannot stand in for either conversation budget.
For self-serve API usage, pay-as-you-go billing (PAYG) is prepaid, with subscription credits consumed first. Existing enabled usage-based subscribers and Enterprise contracts retain overage billing. Voice slots and cloning eligibility still follow the subscription tier; adding a PAYG balance does not buy PVC access. A depleted balance can pause service.
Hume's Audio Decisions billing rate is $0.47 per million input tokens, counting audio and question/context text, with answers unbilled. That buys analysis after access is granted. It is not a cheaper speech-generation rate and cannot be converted into an Octave replacement quote.
Where Dasha fits for a production voice agent
A narration product may need only a synthesis API. A production voice agent also needs telephony or web deployment, integrations and visibility into what happened during a conversation.
At Dasha, we provide a managed production platform for conversational AI products. For technical teams that want managed agent execution around a configurable speech component, we recommend evaluating our runtime alongside the voice-model decision.
Our TTS provider configuration includes ElevenLabs v4 and v4 Turbo. Our phone and web deployment workflows provide the surrounding execution path, while completed-call inspection brings together transcripts, recordings, language-model interactions, tool executions and latency breakdowns.
That is a runtime-and-operations choice. It does not make Hume voice IDs portable, translate EVI behavior automatically or imply that every provider-specific control has an equivalent setting.
For a managed foundation around your replacement speech component, explore our voice AI backend.
