multi-session-dialogue-questions-whose-answers-depend-only-how-utterance CLM-0451On multi-session dialogue questions whose answers depend only on how an utterance was delivered, a cascaded ASR-then-LLM pipeline (Whisper large-v3 into a frontier model) scores below 7B audio-native models on the same clips, and below even a transcript-only baseline, because the transcription step drops delivery cues entirely and adds content errors.
Capability: Understanding speech and audio
Observed on
Synthesized two-speaker evidence sessions (104 clips, TTS with acted-emotion reference conditioning), evidence-only context of roughly 10-15k tokens; only two 7B audio-native model.
Sources
- Measured: Whisper->Opus 4.8 scored 0.254 vs Qwen2-Audio 0.354 and Qwen2.5-Omni 0.412, with a blind text baseline at 0.325 and descriptive text at 0.675; transcript analysis found no bracketed voice events or fillers in any of the 104 transcripts. Cascade and audio arms differ in model scale and in modality, so the comparison does not isolate scale, and the audio is entirely TTS-generated.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For spoken language models answering questions over long dialogue histories, converting the history to transcripts beats keeping it as audio when the answer is a lexical detail (exact slot values across dense turns: mean 71.3% vs 42.6%), but loses when the answer depends on how something was said, where audio history wins for six of seven models and by a larger margin once lexical emotion cues are removed.Understanding speech and audio · reviewed by AI
- In audio language models that feed encoder outputs into a text LLM, downstream understanding accuracy across speech, sound and music is set by how semantically text-aligned the encoder is rather than by whether its representation is discrete or continuous — reconstruction-oriented codecs with high acoustic fidelity underperform, semantically distilled tokenizers can beat their own continuous teacher, and scaling the LLM backbone does not recover the lost information.Understanding speech and audio · reviewed by AI
- When post-training a speech LLM for ASR with GRPO, adding a reward term from a separately pretrained, frozen CTC acoustic judge that scores each sampled transcription against the audio cuts insertion errors on out-of-distribution meeting speech (22-28% fewer on AMI) relative to a schedule-matched text-only -WER reward, without raising word error rate and without needing the judge at inference.Stating false facts confidently · reviewed by AI
- In LLM-based ASR, steering decoding toward acoustic evidence only helps if candidates are first restricted to tokens within a small likelihood margin of the greedy token: unconstrained selection by an acoustic compatibility score removes hallucinated outputs but destroys transcription accuracy through unbounded insertions, while the constrained version removes a large share of hallucinations at near-zero WER/CER cost.Stating false facts confidently · reviewed by AI
- When two audio clips are concatenated at the waveform level into one input, current large audio language models keep transcribing both segments nearly as well as single clips while their accuracy on content, order and location questions about the same two segments drops far below their single-clip attribute-recognition accuracy.Understanding speech and audio · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A cascade whose ASR emits or annotates paralinguistic markers, or a frontier cascade that matches or beats audio-native models on the same delivery-dependent items, would show the loss is not architectural.