oruk/oruk-bench-leaderboard

oruk-bench leaderboard

Results for 64 speech emotion recognition systems measured on one held-out multilingual evaluation with a single scoring implementation: open checkpoints, closed APIs, audio LLMs, and text-only baselines, all on the same protocol.

This dataset is the results table, not the audio. The evaluation clips are assembled from several emotional-speech corpora whose licences differ, so they are not redistributable; the benchmark card documents provenance and how to reconstruct an equivalent set. What is published here is the part people want to cite and compare against.

The evaluation

• 64,384 held-out clips, roughly 20 languages, a mix of acted, elicited and spontaneous speech.
• Seven classes: anger, happiness, sadness, fear, disgust, surprise, neutral.
• Single-label classification. 16 kHz mono, truncated to the first 16 seconds. Systems with other native taxonomies are mapped through a published alias table.
• Primary metric is macro-F1, which is what to read when classes are imbalanced. Accuracy is reported alongside it.

Fields

Field	Meaning
system	Model or product name as evaluated
family	ours, open, API, audio LLM, or text only
accuracy	Percent, 0–100
macro_f1	Macro-averaged F1, 0–1. The primary metric
scored_on	full_set (all 64,384 clips) or subsample_5k
is_oruk	Whether the row is an oruk model
note	Dated snapshot caveats, where a row must not be read as current-product coverage

Composition: 29 open, 18 audio LLM, 11 API, 5 text-only, 1 oruk. Thirty systems were scored on the full set and thirty-four on the subsample.

Two caveats that change how the table reads

oruk Spectra is trained in-distribution. Every other entrant is zero-shot cross-corpus. Spectra leads at 77.6% accuracy and 0.810 macro-F1, against 68.7% and 0.683 for the strongest open baselines in the emotion2vec+ family. That gap is not a like-for-like comparison and should not be cited as one. It is published in the same table anyway, flagged by is_oruk, because omitting it would be worse.

full_set and subsample_5k are not interchangeable columns. Closed APIs and local audio LLMs are scored on a fixed label-stratified 5,000-clip subsample, seed 0, because running them on 64,384 clips is prohibitive. Rescoring open models on the same subsample shifts scores by less than two points, which is why the columns sit side by side — but the distinction is recorded per row rather than smoothed away.

What the results show

The spread is very wide. Macro-F1 runs from 0.810 at the top to 0.079 at the bottom, and a substantial number of well-known systems land under 0.40. Cross-corpus multilingual speech emotion recognition is much harder than single-corpus reported figures suggest, and the headline numbers published for many of these systems do not survive contact with held-out audio from a different distribution.

Two rows worth pulling out for anyone working on compressed models: HuBERT-large SUPERB scores 40.9% / 0.249, while DistilHuBERT SUPERB scores 29.5% / 0.186. Distillation is not close to free for this task, where the discriminative signal is prosodic rather than lexical.

The five text-only baselines exist to answer the obvious objection: how much of this is recoverable from the words alone, without hearing the audio?

Intended use, and what this must not be used for

Appropriate: comparing architectures and training recipes on a common protocol, tracking progress on cross-corpus multilingual emotion recognition, auditing capability claims, and studying failure modes.

The labels are perception-level annotations of vocal expression. A high score means a system matches how human raters perceive a clip — not that it can determine what anyone actually feels.

This benchmark must not be used to develop, market or validate systems that infer the emotions of natural persons in workplace or education settings, which EU AI Act Article 5(1)(f) has prohibited since 2 February 2025. Recital 44 of the same regulation sets out the scientific objections to the category, and the spread in this table is consistent with them. Full scope in INTENDED_USE.md.

Citation

@misc{orukbench,
  title        = {oruk-bench: a held-out multilingual benchmark for speech emotion recognition},
  author       = {{Oruk AI}},
  year         = {2026},
  howpublished = {\url{https://github.com/Oruk-AI/oruk-bench}},
  note         = {64 systems evaluated on 64,384 held-out clips across ~20 languages}
}

Links

• Interactive leaderboard: oruk.ai/benchmarks
• Methodology: oruk.ai/benchmarks/methodology
• Code, benchmark card and governance: Oruk-AI/oruk-bench
• Try the top-scoring model: Space