Speech clips used by ConductCognition tests Version 1.5.0, made on 2026-11-10 by scripts/speech-kit/make_clips.py of the ConductCognition source. The manifest (manifest.json) lists every clip with its text, voice, engine, settings, level and SHA-256 hash, and the version of the clips that added each set (sets). Synthetic voices. The digits, sentences, words and phrases, and the monitoring sentences and restoration words, are our own texts, spoken by open text-to-speech models. No voice is a copy made to sound like a named person. For each language, the voice, its terms and its training data's terms: - en: Piper en_US-ljspeech-high (American English; a woman's voice, by its model card). piper-tts 1.8.0 (GPL-3.0-or-later) with onnxruntime 1.31.1 (MIT). Model: https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/ljspeech/high/en_US-ljspeech-high.onnx Terms: voice model MIT (huggingface.co/rhasspy/piper-voices); the engine's GPL does not cover its output Training data: trained from scratch on the LJ Speech dataset, public domain (keithito.com/LJ-Speech-Dataset) - es: Piper es_ES-carlfm-x_low (Spanish of Spain). piper-tts 1.8.0 (GPL-3.0-or-later) with onnxruntime 1.31.1 (MIT). Model: https://huggingface.co/rhasspy/piper-voices/resolve/main/es/es_ES/carlfm/x_low/es_ES-carlfm-x_low.onnx Terms: voice model MIT (huggingface.co/rhasspy/piper-voices); the engine's GPL does not cover its output Training data: trained from scratch on carlfm01's speech dataset, public domain (github.com/carlfm01/my-speech-datasets) - es, the digits: Coqui tts_models/es/css10/vits (Spanish). coqui-tts 0.28.0 (MPL-2.0) with torch 2.8.0 (BSD-3-Clause). Model: https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--es--css10--vits.zip Terms: BSD-3-Clause (Coqui's model list; Neon Gecko's model, huggingface.co/neongeckocom/tts-vits-css10-es) Training data: CSS10 Spanish: LibriVox readings by Tux, public domain; CSS10's repository is Apache 2.0 (github.com/Kyubyong/css10); the model card is empty, and its configuration names css10_es as its only data - fr: Coqui tts_models/fr/css10/vits (French). coqui-tts 0.28.0 (MPL-2.0) with torch 2.8.0 (BSD-3-Clause). Model: https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--fr--css10--vits.zip Terms: BSD-3-Clause (Coqui's model list; Neon Gecko's model, huggingface.co/neongeckocom/tts-vits-css10-fr) Training data: CSS10 French: LibriVox readings by Gilles G. Le Blanc, public domain; CSS10's repository is Apache 2.0 (github.com/Kyubyong/css10); the model card is empty, and its configuration names css10_fr as its only data - de: Coqui tts_models/de/thorsten/vits (German). coqui-tts 0.28.0 (MPL-2.0) with torch 2.8.0 (BSD-3-Clause). Model: https://github.com/coqui-ai/TTS/releases/download/v0.7.0_models/tts_models--de--thorsten--vits.zip Terms: Apache 2.0 (Coqui's model list, TTS/.models.json) Training data: the Thorsten-Voice dataset, neutral, December 2021, CC0 (github.com/thorstenMueller/Thorsten-Voice) - it: Piper it_IT-riccardo-x_low (Italian). piper-tts 1.8.0 (GPL-3.0-or-later) with onnxruntime 1.31.1 (MIT). Model: https://huggingface.co/rhasspy/piper-voices/resolve/main/it/it_IT/riccardo/x_low/it_IT-riccardo-x_low.onnx Terms: voice model MIT (huggingface.co/rhasspy/piper-voices); the engine's GPL does not cover its output Training data: trained from scratch on the M-AILABS Speech Dataset, Italian (Riccardo Fasol's LibriVox readings): LibriVox audio and texts in the public domain, the dataset's own license allowing any commercial use with its notice kept (github.com/imdatceleste/m-ailabs-dataset) - pt: Coqui tts_models/pt/cv/vits (Portuguese; one speaker, "male" in its speaker file). coqui-tts 0.28.0 (MPL-2.0) with torch 2.8.0 (BSD-3-Clause). Model: https://github.com/coqui-ai/TTS/releases/download/v0.8.0_models/tts_models--pt--cv--vits.zip Terms: BSD-3-Clause (Coqui's model list; Neon Gecko's model, huggingface.co/neongeckocom/tts-vits-cv-pt) Training data: Common Voice Portuguese, one speaker, CC0 (its configuration names pt_CV/clean.tsv; the model card is empty) - zh: PaddleSpeech fastspeech2_aishell3 with hifigan_aishell3, speaker 6 (AISHELL-3 speaker SSB0033; Mandarin). PaddleSpeech FastSpeech2 and HiFiGAN, ONNX (Apache-2.0), run with onnxruntime 1.31.1 (MIT); front end scripts/speech-kit/zh_front.py with pypinyin 0.55.0 (MIT). Model: https://paddlespeech.cdn.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_aishell3_onnx_1.1.0.zip and .../hifigan/hifigan_aishell3_onnx_1.1.0.zip Terms: Apache-2.0 (github.com/PaddlePaddle/PaddleSpeech) Training data: AISHELL-3, Apache 2.0 (openslr.org/93; huggingface.co/datasets/AISHELL/AISHELL-3) - ja: Coqui tts_models/ja/kokoro/tacotron2-DDC with its kokoro HiFiGAN (Japanese). coqui-tts 0.28.0 (MPL-2.0) with torch 2.8.0 (BSD-3-Clause). Model: https://huggingface.co/tts-hub/tacotron2_ddc-ja-kokoro Terms: Apache 2.0 (model card and Coqui's model list) Training data: the Kokoro Speech Dataset, public domain: LibriVox readings by ekzemplaro of Aozora Bunko texts (github.com/kaiidams/Kokoro-Speech-Dataset) Dichotic syllables (dichotic/): ba, da, ga, pa, ta, ka, made by formant synthesis in scripts/speech-kit/dsp.py, after Klatt, D. H. (1980), Software for a cascade/parallel formant synthesizer, Journal of the Acoustical Society of America 67(3), 971-995. No voice model or recording. Speech-shaped noise (noise/): made in scripts/speech-kit/dsp.py with the long-term spectrum of each language's ten digits. No recording. Vocal emotion clips (emotion/): a subset of CREMA-D, the Crowd-sourced Emotional Multimodal Actors Dataset. Cao H, Cooper DG, Keutmann MK, Gur RC, Nenkova A, Verma R. CREMA-D: Crowd-sourced Emotional Multimodal Actors Dataset. IEEE Transactions on Affective Computing. 2014;5(4):377-390. doi:10.1109/TAFFC.2014.2336244 Source: https://github.com/CheyneyComputerScience/CREMA-D (AudioWAV, and the voice-only ratings in processedResults/tabulatedVotes.csv). CREMA-D is made available under the Open Database License: http://opendatacommons.org/licenses/odbl/1.0/. Any rights in individual contents of the database are licensed under the Database Contents License: http://opendatacommons.org/licenses/dbcl/1.0/. This subset, its clips and their rating fields in manifest.json, is a derivative of CREMA-D and is made available under the same Open Database License 1.0, its contents under the Database Contents License 1.0. Changes: the clips were chosen by the rule in scripts/speech-kit/items.json, high-pass filtered at 60 Hz, trimmed to 20 ms before the speech and 50 ms after it, set to one active speech level, and saved as FLAC. Added in version 1.1.0, on 2026-11-05: - A /ba/ to /da/ continuum (continuum/): 9 steps made by the same formant synthesis as the dichotic syllables, in scripts/speech-kit/dsp.py; step 1 is the dichotic ba and the last step its da. No voice model or recording. - Speech-to-song phrases (phrases/): ten short statements per language, our own texts, spoken by each language's voice above. No recording. Added in version 1.2.0, on 2026-11-06: - Phoneme monitoring sentences (monitoring/en/): English sentences, our own texts, spoken by the English voice above. No recording. - Phonemic restoration words (restoration//): 36 words per language, our own, each said alone by its language's voice above, and each also with Gaussian white noise made by scripts/speech-kit/dsp.py laid over its s sound (-added) or put in its place (-replaced). No recording. Added in version 1.3.0, on 2026-11-06: - Keyword sentences (keyword-sentences//): 59 everyday sentences per language, our own texts, each with three key words, spoken by that language's voice above. No recording. - Sentence noise (sentence-noise/): made in scripts/speech-kit/dsp.py with the long-term spectrum of each language's keyword sentences. No recording. Added in version 1.4.0, on 2026-11-09: - Mandarin lexical tones (lexical-tone/zh/): five syllables in the four dictionary tones, our selection, spoken by the Chinese voice above. The syllable mi is for practice only. No recording. - Dictation (dictation//): 27 everyday words per language, our selection, spoken by each language's voice above; the last three are for practice only. No recording. Added in version 1.5.0, on 2026-11-10: - Voice identity (voice-identity//): original everyday sentences, synthesized with eSpeak NG 1.52.0. Each of en, de, es, fr, it, pt, zh (cmn) and ja uses m1, m2, f1, f2; practice uses m3 and f3. These are rule-based formant voices with no recordings, neural model or imitated person. Generator and voice definitions: GPL-3.0-or-later; commercial use is allowed. Section 2 excludes ordinary audio output from the program license. We ship generated speech, no executable or voice-definition file. Source: https://github.com/espeak-ng/espeak-ng/tree/1.52.0 Voices: https://github.com/espeak-ng/espeak-ng/tree/1.52.0/espeak-ng-data/voices/!v Terms: https://raw.githubusercontent.com/espeak-ng/espeak-ng/1.52.0/COPYING (read 2026-11-10). The manifest records each definition hash, generator version and settings. All identities speak the same texts within a language; study, recognition and pair judgments use separate sentences. Tests that play these clips: - Dichotic Listening Task: the six dichotic syllables, listed in src/tests/stimuli/dichotic-listening.ts of the ConductCognition source. - Digits-in-Noise Task: the ten digits and the noise of the session's language, all but the Japanese 4, listed in src/tests/stimuli/digits-in-noise.ts of the ConductCognition source; with its sentences setting, the 59 keyword sentences and the sentence noise of the session's language, listed in src/tests/stimuli/digits-in-noise-sentences.ts. - Listening Span Task: the sentences of the session's language, less the known misses and, in Chinese and Japanese, the sentences whose word to remember is not the last thing named, listed in src/tests/stimuli/listening-span.ts of the ConductCognition source. - Dual N-Back Task: the digits 1, 2, 3, 5, 6, 8 and 9 of the session's language, listed in src/tests/stimuli/dual-nback.ts of the ConductCognition source. - Vocal Emotion Task: all 54 emotion clips, 48 in the test and 6 in practice, listed in src/tests/stimuli/vocal-emotion.ts of the ConductCognition source. The test is a Produced Work of the CREMA-D subset above, which stays under the Open Database License 1.0, its contents under the Database Contents License 1.0. - Cross-Modal Priming Task: 78 of the 80 English words, all but cup and bee, listed in src/tests/stimuli/cross-modal-priming/en.ts of the ConductCognition source. - Categorical Perception Task: the nine continuum steps, listed in src/tests/stimuli/categorical-perception.ts of the ConductCognition source. - Speech-to-Song Task: the ten phrases of the session's language, listed in src/tests/stimuli/speech-to-song.ts of the ConductCognition source. - Phoneme Monitoring Task: the 86 monitoring sentences, a session playing 54 of them, listed in src/tests/stimuli/phoneme-monitoring-clips.ts of the ConductCognition source. - Phonemic Restoration Task: the 36 words of the session's language, the demonstration word as it is and every word's two versions, 73 clips, listed in src/tests/stimuli/phonemic-restoration-clips.ts of the ConductCognition source. - Lexical Tone Task: all 20 Mandarin syllable clips, 16 for test pairs and four for practice, listed in src/tests/stimuli/lexical-tone.ts of the ConductCognition source. - Spelling to Dictation Task: the session language's 27 words, 24 for the test and three for practice, listed in src/tests/stimuli/spelling-dictation.ts of the ConductCognition source. Voice Identity Task (voice-identity): voice-identity//, all six synthetic identities.