Read a conversation one passage at a time
Listen to this 14.45-second conversation about grocery prices, then compare the four passages below. The words, timestamps, speaker labels and scores come directly from the saved oruk-resonance response. Playing the audio makes no new analysis request.
The four passages below contain the full returned transcript. Scores are rounded to three decimals; the linked JSON preserves their original precision.
Passage 1
0.185–4.325 s
speaker_0
“I'm sorry. The price of food has gone up so much. I mean...”
- Emotion scores
- disappointed 0.868
- Speaking-style scores
- casual 0.869
Passage 2
4.325–6.065 s
speaker_1
“It really has.”
- Emotion scores
- happy 0.780 · neutral 0.756
- Speaking-style scores
- passionate 0.910
Passage 3
6.065–7.805 s
speaker_0
“What what happened?”
- Emotion scores
- angry 0.903 · frustrated 0.849 · disgusted 0.847
- Speaking-style scores
- irritated 0.905 · impatient 0.826
Passage 4
8.805–14.125 s
speaker_1
“I don't know. Like, honestly, I don't know what happened.”
- Emotion scores
- disgusted 0.845 · frustrated 0.827 · angry 0.815
- Speaking-style scores
- sarcastic 0.949 · passionate 0.923 · irritated 0.910 · skeptical 0.882
Audio: The Agentic Data Company, Open Yap 1K, CC BY 4.0. Oruk excerpted and downmixed the recording. Attribution · File hashes and provenance. This saved response has no inference date or model-build identifier. It is a worked example, not a current accuracy measurement or a customer result.
What the example tells you
The first passage returns disappointed at 0.868. The short reply “It really has” returns both happy and neutral. Listen before deciding whether either annotation is useful for your review task. Several labels can appear together; these scores do not form a probability distribution over the speaker’s feelings.
The third passage returns frustrated at 0.849, and the fourth at 0.827. They belong to different speaker labels. Joining them into one frustration trend would lose that distinction. Keep each score with its passage and speaker; speaker_0 and speaker_1 do not identify customer and agent.
This response contains no word timestamps. Its boundaries support passage review, not a score for every word or sentence. For exact interval replay, filtering and the commands to analyze your own recording, follow the Python call-review walkthrough. Check the label definitions and selection rules before choosing a review threshold.
Where teams apply it
You can use these outputs in several call-review workflows: contact center analytics for queue-wide scoring and QA, voice of customer analytics for trending how customers sound over time, and sales call analysis for coaching against specific moments in a deal.
Published benchmark, with its limits
speech-emotion-bench
77.8% accuracy
Top measured result in our July 2026 release. Open models use the full 64,384 held-out clips; closed/API and audio-LLM systems use a fixed 5,000-clip subset. The highest-scoring non-oruk row is emotion2vec+ seed at 68.7%. The caveat, stated plainly: the oruk entry is trained in-distribution while other systems are evaluated zero-shot. This seven-class result does not establish the current 15-label endpoint’s accuracy on your recordings. Protocol and downloadable results are on the methodology page.
What a response means
A returned emotion either passed a model-specific threshold or was the highest-scoring fallback when none passed. Presence alone is not a confidence test. A score of 0.83 is a model output, not an 83% probability of someone’s feelings. Validate scores on your speakers and recording conditions before setting application rules.
Try it yourself first: run the live analyzer on your own voice, no account needed.
Data handling: Oruk’s inference services discard audio and outputs after responding. Optional speaker diarization retains uploaded audio for up to 48 hours and speaker-label results for up to 24 hours. Oruk does not use customer audio or outputs to train, fine-tune, or evaluate models without your explicit written agreement. Processing details are in the privacy policy.
Plans from $9/month. $0 today, cancel any time.
Exactly what you are buying
The limits below are stated here so you do not discover them after integrating.
- File analysis and live phrase emotion
- Full emotion and speaking-style analysis uses completed English recordings. Use oruk-realtime for live transcript tokens and phrase-level emotion; speaking-style analysis remains in the file API.
- Input
- Resonance accepts prerecorded English audio: WAV, FLAC, MP3, M4A, OGG, or WebM, up to 30 MB and 60 minutes per file.
- Realtime preview
- The separate Realtime preview streams transcription in 32 locales with automatic language detection and phrase-level emotion scores. Full 15-emotion and 16-style analysis uses the English file endpoints.
- File limits and formats
- Up to 30 MB and 60 minutes per file. WAV, FLAC, MP3, M4A, OGG, and WebM are supported.
- Request behavior
- File endpoints return a synchronous JSON response with measured audio duration. Realtime uses WebSocket events for live tokens and phrase-level emotion.
- Pricing
- Monthly plans start at $9. Resonance, Fourier, Realtime, and Proficiency share one audio-minute allowance. Orukeet has a separate allowance; its optional tasks use that allowance at their listed rates. The other-model allowance ranges from 250 minutes on Hobby to 20,000 on Production. Extra usage follows each allowance's published rates. Self-serve plans start with a 7-day free trial. Contact us for Enterprise pricing. Details on the pricing page.
- Data retention and training
- Oruk’s inference services discard audio and outputs after responding. Optional speaker diarization retains uploaded audio for up to 48 hours and speaker-label results for up to 24 hours. Oruk does not use customer audio or outputs to train, fine-tune, or evaluate models without your explicit written agreement.
- What scores are — and are not
- Emotion scores are model annotations selected with model-specific thresholds, not probabilities of inner feelings. They are not intent, truth, medical, employment, or psychological judgments, and must not be the sole basis for consequential decisions. See responsible use.
- Model and version stability
- Models are versioned; Resonance and Fourier are stable, Realtime and Proficiency 1 are in preview. Changes are announced on the changelog.
- Support
- Email access@oruk.ai; the production guide covers retries, limits, and error handling.
Plans from $9/month. $0 today, cancel any time.
Built on an emotion recognition API
Underneath, this is one call to the speech emotion recognition API; this page covers how that signal reads as sentiment on recorded calls, while AI call analytics covers the wider category. Prefer to see it first? Try the captioning demo or read the API docs.
Plans from $9/month. $0 today, cancel any time.
Frequently asked questions
- What is voice call sentiment analysis?
- Voice call sentiment analysis measures the emotion and tone of a spoken call from its audio — not just the words. The oruk Speech API returns a transcript plus selected emotion and speaking-style scores for timed acoustic segments, so you can review how a caller sounded and how that shifted through the conversation.
- Can you analyze sentiment and emotion per sentence in a call?
- Results use time-local acoustic segments, and their boundaries need not match sentences. A segment can contain several sentences. Use its start and end timestamps to locate the returned emotion and speaking-style scores in the recording.
- Can I use this for live agent assist?
- Yes, for live transcript tokens and phrase-level emotion scores through oruk-realtime. The full 15-emotion and 16-speaking-style analysis still runs on a completed English recording for post-call scoring, coaching, and quality review.
- How accurate is emotion scoring on call audio?
- Scores are model annotations selected with model-specific thresholds, not probabilities of a caller’s inner state. Published benchmarks describe dated evaluations; they do not establish accuracy on your calls or the current endpoint. Validate on recordings from your microphones, codecs, accents, and call types.
- How is this different from an emotion recognition API?
- It is the same emotion recognition API. This page covers how that signal applies to recorded calls; the emotion recognition API page documents the raw endpoint and the label space.
Try it free for 7 days
Plans start at $9 a month, with access to every speech model. Add a card, pay nothing today, and cancel any time before the trial ends.