Keep the speech signals your product relies on.
Moving from Hume starts with the behavior your application needs. Identify the dependency, test a replacement on your recordings, and cut over with evidence.
Start with the Hume product you use.
Speech measurement
Keep your own recordings and archived results. Compare the specific labels, timing and failure behavior your application uses, then recalibrate its decisions.
Compare measurement contractsEvaluate speech measurement · optional Oruk account
EVI voice agents
Rebuild the conversation stack with your chosen transcription, language model and voice providers. Oruk can add acoustic expression signals alongside that stack.
Plan an EVI replacementEvaluate Oruk expression signals · optional Oruk account
TTS and Octave
Choose a speech-generation provider and verify voice rights, quality and latency. Oruk does not synthesize speech or make Hume voice IDs portable.
Review voice requirementsInspect the dependency locally.
The offline Python tool reads only the files you specify. It reports recognized Hume calls and native contract mismatches. It makes no network requests and does not include source text, credentials or transcripts in its report.
A recognized import is a clue, not proof of compatibility. Review the report before sharing it. Use Python 3.10 or later on macOS or Linux.
Replace the example requirements with your own. Run against an explicit source file or lockfile; repeat --source for additional files.
python3 hume-migration-assess.py \
--source your-client.py \
--requirements hume-migration-manifest.example.json \
--output assessment.jsonAlready have legacy batch responses? Use the archived-response comparison guide to preserve original scores and check audio identity and time origins.
Replay a batch after your first evaluation
The batch runner checks an explicit manifest of original WAV files offline. Execution sends audio to Oruk and can incur usage charges. It keeps a private resume ledger and never automatically resends an attempted request.
Choose a contract you can test.
These routes have different formats, timing and access requirements. Evaluate them against your application’s outputs and decisions.
ResonanceStable
/v1/audio/analysis
Synchronous file request. English WAV, FLAC, MP3, M4A, OGG or WebM; up to 30 MB and 60 minutes.
Transcript and timed segments with selected scores from 15 emotions and 16 speaking styles. Optional speaker labels are local to the recording.
Selected emotions use thresholds with a highest-score fallback; styles can be empty. Missing labels are not zero. Use /emotions or /affect without a transcript.
Server-side bearer API key and available account usage.
Best starting point for an English legacy batch workflow that needs timed results. Validate the new vocabulary and decision rules.
Resonance documentationSpectra-2Stable
/v1/audio/analysis
File request or WebSocket upload, followed by finish. Automatic transcription in 25 languages. Mono 16 kHz WAV, finalized FLAC, PCM16 or float32le; 45 ms–60 seconds, up to 4 MiB.
Final transcript and all 15 emotion and 16 style scores for the whole recording. No timed segments, word timestamps, speaker labels or partial transcript.
Language coverage describes transcription support, not a guarantee of equal expression accuracy by language. Scores are independent and do not sum to one.
Server-side bearer API key; account access and usage checks apply. Prepared browser connections use short-lived tickets issued by your server.
For short multilingual recordings. Streaming uploads audio as it arrives; analysis returns after finish. This is not the Realtime phrase-event protocol.
Spectra-2 documentationResonance 2Preview
/v1/audio/resonance-2
File request or explicit sample-range WebSocket commits. Files: 0.1–120 seconds, up to 30 MiB. WebSocket: mono 16 kHz float32le PCM.
31 category scores, 6 signed axes and 19 independent scores for each requested range. No transcript, diarization or model-generated timestamps.
Selected labels may be empty; no forced top label. Signed expression axes are not validated valence/arousal coordinates. Keep model and calibration revisions with results.
Server-side bearer API key and approved organization model access; a paid plan alone does not grant approval.
For acoustic expression analysis beside your existing transcription/agent stack. Measure full request latency and validate language/domain performance.
Resonance 2 documentationRealtimePreview
/v1/realtime
WebSocket transcript deltas and independent phrase events. Mono 16 kHz PCM16; transcription in 32 locales. Sessions are limited to 10 minutes.
Live transcript deltas and independently timed phrase emotion events. Read the variable-length label/score arrays; do not hard-code seven labels.
A phrase result can arrive after final text or fail independently. Join by request/turn/phrase identity. Locale support is not per-language emotion validation.
Authenticated WebSocket handshake. Keep permanent keys server-side; do not put keys in URLs or public browser bundles.
For live transcripts with phrase-level signals. Its PCM format, events and turn lifecycle differ from Resonance 2 and Spectra-2.
Realtime documentationWhat this migration does not provide
- Hume-compatible scores, thresholds or a 48/600-dimension taxonomy
- EVI session, tool, voice or audio-generation compatibility
- First-party text-to-speech or automatic voice portability
- Face/image analysis, text sentiment or a dedicated vocal-burst API
Budget for your selected Oruk model and any separate transcription, language-model or voice providers. See current plans and trial terms.
Bring the workflow.
We’ll help scope the move.
Tell us which outputs matter, where the application runs, and when you need to switch. We’ll review the fit and identify the tests required before a cutover.
Keep recordings and exports in your own storage. Start with the customer export checklist.