Sarcasm detection from voice

The honest framing: delivery, not intent

Sarcasm is ultimately a claim about what a speaker meant, and no model can verify intent from sound alone. What audio does carry is the performance of sarcasm: speakers mark it acoustically so listeners catch it. oruk scores that marking with its sarcastic speaking-style label — one of 16 — and states the limit plainly in its capabilities and scope: no intent detection, ever. Treat a high sarcastic score as “this sounded sarcastic,” a strong prior for a human reviewer, not a verdict.

What the research supports

Sarcasm is not carried by one magic acoustic feature. The original MUStARD dataset paper found that multimodal and conversational context improved over isolated, single-channel baselines. A recent systematic review of sarcasm recognition reaches the same practical conclusion: words, vocal delivery, and context can disagree, and robust systems need to represent that disagreement. oruk exposes the acoustic evidence as a style score; your application keeps the conversational context and the final decision.

The acoustic cues

Exaggerated prosody

Sarcasm often over-performs: stretched vowels, sing-song contours, theatrical stress ("Oh, greeeat").

Flattened delivery

The opposite pattern also signals it: deliberately affectless, deadpan phrasing where enthusiasm would be expected.

Tone–word mismatch

Positive words carried by negative or flat prosody. The contradiction between channels is the cue humans use most.

Timing

Delayed onsets, drawn-out syllables, and pointed pauses that frame the utterance as performance.

Scoring sarcastic delivery via API

One request against a prerecorded English audio file scores a vocabulary of 16 speaking styles and returns labels that pass their model-specific thresholds. An empty styles array is valid.

curl https://speech-api.oruk.ai/v1/audio/styles \
  -H "Authorization: Bearer $ORUK_API_KEY" \
  -F file=@clip.wav -F model=oruk-resonance

{
  "object": "audio.styles",
  "model": "oruk-resonance",
  "styles": [
    { "label": "sarcastic", "score": 0.81 },
    { "label": "deadpan", "score": 0.64 }
  ],
  "usage": { "audio_seconds": 6.2, "billable_seconds": 6.2 }
}

Pair with /v1/audio/analysis to get the transcript and emotion labels in the same response — useful for surfacing tone–word mismatches, where positive words meet a sarcastic or deadpan delivery.

Where teams use it

• QA review queues: flag calls where “satisfied” words carried sarcastic delivery, so sentiment dashboards stop counting them as wins.
• Research coding: pre-annotate interview audio with style labels for human coders to confirm, instead of listening to everything.
• Agent evaluation: measure whether voice-agent conversations end with sincere versus sarcastic acknowledgments.

In every case the label routes attention; a person makes the judgment. See responsible use.

FAQ

Can AI detect sarcasm in speech?

Partially, and the framing matters. Acoustic models can detect sarcastic-sounding delivery — the exaggerated or flattened prosody speakers use to signal sarcasm — with useful accuracy. No system can verify a speaker’s actual intent. oruk scores "sarcastic" as one of 16 speaking-style labels selected with model-specific thresholds and is explicit that this describes how the speech sounds.

What is the difference between sarcasm detection and sarcastic-style detection?

Sarcasm detection implies knowing what the speaker meant — an intent claim that audio alone cannot support. Sarcastic-style detection scores the acoustic delivery pattern associated with sarcasm. oruk deliberately offers the second and not the first; the capabilities page states that API v1 does no intent detection.

Why do text-only models miss sarcasm?

Because "great, thanks" transcribes identically whether sincere or biting. On speech-emotion-bench, text-only LLMs working from transcripts scored 37–41% seven-class emotion accuracy — the acoustic channel carries the signal the words leave out.

How do I get sarcasm scores from the oruk API?

POST an audio file to /v1/audio/styles (or /v1/audio/analysis for the transcript, emotion, and style together). The model scores a vocabulary of 16 style labels, including sarcastic and deadpan, and returns those selected with model-specific thresholds. The styles array can be empty. Plans cover the speech model lineup. New organizations need separate approval for Resonance 2; an active trial or plan does not by itself grant access. Hobby is $9/month, Builder $49/month, and Production $199/month. Spectra-2, Resonance 1, Resonance 2, Fourier, Realtime, and Proficiency share 250, 2,500, or 20,000 monthly audio minutes, respectively, including supported diarization. Each plan also includes a separate Orukeet allowance, starting at up to 20,000 transcription minutes on Hobby. Orukeet uses $0.00045/minute ($0.027/hour); optional tasks draw from its allowance at their published rates.