Skip to content

Try Resonance 2 early.

Tell us a little about yourself and what you’re building.

A brief description of your use case.

We’ll use these details to review your application and contact you about access. Privacy policy

Get started free

Explore oruk

Plan an EVI replacement around your application

EVI combines a conversation runtime, language model, voice generation and expression signals. Replacing it means selecting and testing those parts together. Oruk supplies speech analysis; your chosen providers supply the rest of the conversation stack.

Reviewed October 3, 2026. This guide defines a migration checklist, not a tested drop-in EVI replacement.

Export first, then verify the archive.

Use the export methods available to your Hume account. Availability, retention and recording permissions vary; do not assume every artifact can be recovered. Store exports in an encrypted location you control.

  1. Inventory EVI configuration versions, system prompts, language models, tool schemas, voice identifiers, prosody settings, webhooks and client SDK versions. Record the active configuration each application actually uses.
  2. Export permitted chat history, events and audio where available. Record which requested artifacts were unavailable. Keep source formats and timestamps; EVI millisecond offsets need explicit conversion when compared with second-based audio results.
  3. Inventory voice metadata, licensed source recordings and permission records separately. A voice identifier does not transfer a voice model or grant cloning rights.
  4. Create a manifest with artifact type, configuration version, count, byte size and SHA-256 checksum. Keep customer identifiers and transcripts out of public issue reports.
  5. Restore a small permitted sample from the archive, verify checksums and confirm that an application owner can read it. Record the missing data before shutting down the old workflow.

Keep API credentials out of exports and source control. Issue fresh secrets for the new providers through your normal secret manager; revoke old credentials after the supported migration and archive checks are complete.

Assign each responsibility in the new stack.

Audio transport and turn management

Preserve your microphone or telephony path. Choose a supported orchestration framework and explicitly test endpointing, interruption, cancellation and reconnection.

Transcription

Choose the language coverage, transcript timing and streaming behavior your application needs. Preserve your existing STT provider if it already meets those requirements.

Language model and tools

Port prompts and tool schemas to a supported model. Validate tool authorization on your server; translate history and roles deliberately.

Speech generation

Use a separate TTS provider with the voices, rights, streaming behavior and quotas you need. Oruk does not generate speech.

Optional acoustic expression

Add Oruk beside the audio stream when a validated signal improves the workflow. Bind results to a turn and an audio range, expire late signals, and continue the conversation when analysis fails.

Do not map Hume scores directly to Oruk thresholds. Compare against outcomes or annotations relevant to your customer, keep a separate test set, and allow the policy to abstain when the evidence is weak. Expression scores do not establish a person’s inner state.

Treat voice replacement as its own decision.

For TTS-only or Octave applications, begin with a speech-generation provider. Evaluate pronunciation, languages, first-audio latency, cancellation, rate limits, commercial terms and your rights to use the voice. Re-create a voice only when the people and licensors involved have authorized that use.

Test actual streaming audio and interruption behavior in your transport. An existing provider voice ID is not portable, and a similar voice is not a guaranteed match. Oruk expression analysis is optional and adds its own usage cost.

Make the release test a real conversation.

  • Run a greeting, at least two turns and a permitted tool call. Verify roles, history, tool results and the next spoken response.
  • Interrupt speech while audio is playing. Confirm that canceled audio stops and stale text, tool results and expression signals do not enter the next turn.
  • Disconnect and reconnect. Preserve only intended history, prevent duplicate tool execution and account for requests with uncertain billing completion before retrying.
  • Test silence, long speech, supported languages, provider timeouts, expired credentials, rate limits and analysis failures. A missing expression signal must leave ordinary conversation usable.
  • Measure audio-to-first-response latency, end-of-turn latency, concurrency and total per-session cost with the chosen provider versions. Measure at expected peak load before promising capacity.

Cut over with a fallback that will still exist.

Start with an explicitly opted-in pilot, then expand only after its quality, cost and reliability checks pass. Pin configuration and model revisions, document who can stop rollout, and rehearse the stop. After Hume’s cutoff, rollback must use a tested non-Hume route or a graceful pause.

Before retiring the old path, verify the archive, reconcile usage, revoke unused credentials and record any data or features that were intentionally not migrated.