--- pretty_name: Decider Audio – CLadder (spoken) language: - en license: other license_name: mixed-see-card task_categories: - audio-classification - question-answering size_categories: - 10K Options: A: yes B: no Answer with the option key. ``` The fixed preamble of every CLadder prompt ("Imagine a self-contained, hypothetical world with only the following conditions, and without any unmentioned factors or causal relationships") is part of the text instruction, not of the audio. `question_text` (the original text including the preamble) and `spoken_text` are for quality control and baselines. **Do not give them to an audio model as input.** ## Listening difficulty - The questions are long and full of numbers. 4,336 clips exceed 400 characters and were rendered in segments joined with 180 ms of silence. A listener has to keep several percentages in memory, so the spoken version is harder than the text version in a way that is not about causal reasoning. - About 38 % of the questions belong to stories `nonsense0` to `nonsense9`, whose variables are invented words ("Ricks has a direct effect on zuff and sevue"). Whisper spells such words inconsistently. The mean WER is 0.2275 for these stories and 0.0207 for the others, mostly because of spelling and not missing speech. The templated question texts are partly ungrammatical in the source. ## How it is made 1. **Text.** CLadder v1 (`causalNLP/cladder`, pinned commit, `cladder-v1.zip`, SHA-256 verified). The file `cladder-v1-q-balanced.json` gives all 10,112 questions, one to one. The world description comes from `cladder-v1-meta-models.json` (`background`, joined by `model_id`). 2. **Spoken text.** The background without the fixed preamble, the given probabilities and the question. 3. **Model and voice assignment.** Each question gets one TTS model, balanced (±1) and assigned without looking at the label. Questions with a number of three or more digits go preferably to `omni`, `qwen` and `chatterbox` while those are below their share. The voice is picked by a seeded hash from the test-split voices. 4. **Audio.** 24 kHz mono, 16-bit FLAC. Texts over 400 characters are split at whitespace into segments joined with 180 ms of silence. 5. **Re-rendering, never removal.** Every clip was transcribed with `Systran/faster-whisper-base.en`. A clip fails if the ASR word error rate is above 0.5, if it has more than 30 characters per second, or if it has extra words amounting to 3 or 10 % of the text's words, whichever is larger. A failed clip was rendered again with another voice and seed on the same model, up to 5 renders (461 needed that). 0 clips still failed and keep their best render with `passes_checks = false`. A second repair pass rendered 197 clips again. 37 of them still failed a check and 160 passed the word check but leave out part of their text. A clip got up to 10 more attempts on its own model. Clips on Supertonic and clips that a model could not render at all then moved to another model for up to 3 more. 22 clips now come from another model, which `tts_model` shows. A clip now also passes when its only problem is the word error rate or extra words and its transcript sounds like the text. This check compares phonemes after numbers and units are spelled out, so "topup" and "top up" or "1960" and "nineteen sixty" match. A clip that leaves out part of its text does not pass this way. A truncated clip is replaced only when the new render leaves out less. ### TTS models | TTS model | Checkpoint | Voice mode | Clips | Hours | Mean WER | WER ≤ 0.2 | Re-rendered | Failing | |---|---|---|---:|---:|---:|---:|---:|---:| | `chatterbox` | [ResembleAI/chatterbox](https://huggingface.co/ResembleAI/chatterbox) | clone | 1,683 | 12.9 | 0.1026 | 75.9% | 80 | 0 | | `omni` | [k2-fsa/OmniVoice](https://huggingface.co/k2-fsa/OmniVoice) | clone | 1,694 | 13.9 | 0.1008 | 75.3% | 72 | 0 | | `pocket` | [kyutai/pocket-tts](https://huggingface.co/kyutai/pocket-tts) | clone | 1,690 | 13.9 | 0.0954 | 75.4% | 52 | 0 | | `qwen` | [Qwen/Qwen3-TTS-12Hz-0.6B-Base](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base) | clone | 1,688 | 13.7 | 0.1036 | 73.1% | 189 | 0 | | `sopro` | [samuel-vitorino/sopro-v2-turbo](https://huggingface.co/samuel-vitorino/sopro-v2-turbo) | clone | 1,688 | 14.7 | 0.0979 | 73.6% | 53 | 0 | | `supertonic` | [Supertone/supertonic-3](https://huggingface.co/Supertone/supertonic-3) | preset | 1,669 | 13.9 | 0.0952 | 76.5% | 15 | 0 | OmniVoice, Chatterbox and Qwen are the original PyTorch checkpoints. ## Checks The mean ASR word error rate is 0.0993, and 75.0% of the clips have WER ≤ 0.2. ASR agreement shows that the words were spoken, not that the voice sounds natural. No clip was reviewed by listening. A wrongly spoken percentage barely changes the WER of a long question, so the checks do not catch every number error. ## License and attribution - **Text.** CLadder (Jin et al., 2023), MIT ([repository](https://github.com/causalNLP/cladder)). - **Speech.** OmniVoice weights are reported as CC BY-NC 4.0 (verify on `k2-fsa/OmniVoice`). Qwen3-TTS is Apache-2.0, Sopro is Apache-2.0, Supertonic is OpenRAIL (use restrictions), Pocket TTS is CC BY 4.0 (attribution) and Chatterbox is MIT. - **Reference voices.** Private collection, not redistributed. The dataset is marked `license: other`. Do not use it commercially. ```bibtex @inproceedings{jin2023cladder, title = {CLadder: Assessing Causal Reasoning in Language Models}, author = {Jin, Zhijing and Chen, Yuen and Leeb, Felix and Gresele, Luigi and Kamal, Ojasv and Lyu, Zhiheng and Blin, Kevin and Gonzalez Adauto, Fernando and Kleiman-Weiner, Max and Sachan, Mrinmaya and Sch{\"o}lkopf, Bernhard}, booktitle = {Advances in Neural Information Processing Systems}, year = {2023} } ```