Marimba: A System One Model for Speech
Read the original on arXiv Computation and Language →The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The Flow has not summarised this story yet — read it at arXiv Computation and Language.
Qwen-Audio-Jev extends Qwen-Audio-3.1-ASR-Flash-Next with a dedicated decision head that unifies selection, binary judgment, and scoring into three primitives: CHOICE, NOUL, and SCORE. The model delivers competitive accuracies and error rates across 18 labeled configurations, achieving 78.70% on MMAU-mini, 84.06% on MMSU, 6.50% equal error rate on VoxCeleb1-O, and a 0.1062 normalized MAE on NISQA. Compared to its predecessor, Qwen-Audio-Jev offers roughly a 20‑fold inference speedup, enabling efficient high‑throughput audio decision‑making for tasks such as large‑scale screening, speaker verification, and automated quality assessment.
The paper presents a taxonomy of errors in streaming dictation ASR systems, moving beyond traditional word error rate (WER) metrics. It evaluates eleven ASR systems on a diverse test set of 106 scripted English dictations, 20 clips each in eleven other languages/dialects, and 400 bilingual recordings, categorizing errors into recognition, output formatting, language handling, and audio drop‑out. The study finds that systems with similar WER can fail in markedly different ways—e.g., some miss a large proportion of whispered words or fail to capture non‑English content—highlighting the importance of error type analysis for real‑time dictation. whyItMatters":"The findings reveal that conventional WER alone is insufficient for assessing real‑time dictation performance, as it masks critical failure modes that directly affect user experience."
The study evaluates conversational speech models by placing them at 427 pauses and 482 turn ends from 176 unscripted conversations, scoring their responses on timing, latency, length, wording, and recall of recent dialogue. Hosted APIs (GPT Realtime, Gemini 3.1 Flash Live, Grok Voice) respond at almost every pause, with latency ranging from 1–4 s for silence-based endpointers and about 8 s for semantic turn detection, while full‑duplex models (PersonaPlex, Moshi) remain silent more often and can interrupt within human reaction time. Recall rates vary: hosted systems remember 41–60 % of facts under a Claude Opus 5.5 judge and 19–26 % with a strict word matcher; full‑duplex models recall 17–22 % and 12–15 % respectively.
The paper presents a new evaluation protocol for post‑training quantization of speech language models that separates lexical output, transcript‑insufficient endpoints, and packed implementations. In a Qwen2‑Audio case study, a 6‑bit allocation selected for translation improves chrF scores but degrades emotion recognition, while uniform and front‑layer controls perform better on emotion tasks. Similar patterns hold at 7 bits, and a 4‑bit study shows consistent emotion deficits across all low‑bit allocations, with no advantage for the selected scheme. The study highlights a precision‑dependent mismatch between lexical output, waveform‑dependent behavior, and nominal precision, without claiming a general failure of low‑bit models or a deployment benefit for the selected allocation.
The study evaluates Qwen2‑Audio‑7B‑Instruct quantization at mixed 4/8‑bit settings on English‑to‑German translation and emotion recognition tasks. BLEU and chrF scores relative to FP16 show no significant difference, yet absolute accuracy on RAVDESS drops by 3.71% for 6‑bit and 1.17% for 7‑bit, with the 6‑bit interval statistically significant. The results demonstrate that text‑output metrics alone cannot confirm performance preservation on speech tasks where labels cannot be inferred from transcripts.
arXiv:2609.14817v1 Announce Type: new Abstract: In Yor\`ub\'a, pitch alone separates \d{o}k\d{o} (husband, Mid), \d{o}k\d{\`o} (vehicle, Low), and \d{o}k\d{\'o} (hoe, High) -- the diacritics ARE the...