---

name: audio-text-metrics description: "Use TorchMetrics audio/speech quality metrics and no-download text metrics safely." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

audio-text-metrics

Use this sub-skill when a task needs TorchMetrics audio/speech quality scores or text scores that can run without loading pretrained models or downloading datasets.

Route here

• Intrusive audio quality: SNR, SI-SNR, SDR, SI-SDR, source-aggregated SDR, complex SI-SNR, PESQ, and STOI.
• Speech separation matching: permutation invariant training (PIT) over waveform or source-separation metrics.
• Non-intrusive classical audio metric: SRMR, when its local optional packages are installed.
• Decoded-text metrics: BLEU, SacreBLEU, ChrF/ChrF++, ROUGE without model embeddings, CER, WER, MER, WIL, WIP, TER, EditDistance, and SQuAD.
• Logit metric: Perplexity from language-model logits or unnormalized scores.

Route elsewhere

• Read ../core-api/SKILL.md for Metric lifecycle, update/compute/reset, device placement, distributed synchronization, Lightning logging, and custom metrics.
• Read ../collections-wrappers-plotting/SKILL.md for MetricCollection, wrappers, trackers, and plotting mechanics.
• Read ../model-based-metrics/SKILL.md for BERTScore, InfoLM, CLIP or embedding metrics, DNSMOS, NISQA, and any metric that needs pretrained weights, caches, or network/cache planning.

Read and run

• Read references/audio-text-api.md when choosing metric classes/functions, input shapes, target nesting, optional dependencies, and higher/lower-is-better interpretation.
• Read references/audio-text-workflows.md when scoring generated waveforms, speech separation outputs, ASR transcripts, summaries, translations, SQuAD answers, or language-model logits.
• Read references/troubleshooting.md when imports fail, PESQ/STOI reject audio, ROUGE/SacreBLEU tokenization surprises occur, or perplexity looks wrong.
• Run scripts/audio_text_metric_smoke.py to perform a deterministic no-network smoke check for installed TorchMetrics audio/text metrics; use --audio, --text, or --all.

Safe defaults

1. Treat preds as the model output and target as the reference/ground truth.
2. For waveform-pair metrics, pass float tensors with identical shape and time along the last dimension.
3. For text overlap/error metrics, pass already-decoded strings; do not pass token IDs or logits.
4. For Perplexity, pass float logits shaped [batch, sequence, vocabulary] and integer targets shaped [batch, sequence]; TorchMetrics does not shift language-model outputs for you.
5. Avoid network-sensitive metric choices unless the task explicitly asks for them and the environment/cache is approved.