Global TIMIT Tianjin LDC2026S13 December 18, 2025 Linguistic Data Consortium 1. Overview =========== Global TIMIT Tianjin was developed by Linguistic Data Consortium and Tianjin University and consists of read speech, transcripts, and forced alignments from a set of 50 bilingual Tianjin/English speakers. It is one of several corpora from the Global TIMIT project, which aims to create a series of corpora in a variety of languages with a similar set of key features of the original TIMIT Acoustic-Phonetic Continuous Speech Corpus (LDC93S1): - A large number of fluently-read sentences, containing a representative sample of phonetic, lexical, syntactic, semantic, and pragmatic patterns - A relatively large number of speakers - Time-aligned lexical and phonetic transcription of all utterances - Some sentences read by all speakers (calibration), others read by a few speakers (shared), and others read by just one speaker (unique) Each speaker is recorded reading four different list of sentences: - Simple Chinese -- 120 sentences selected to represent everyday Chinese - Chinese Gigaword -- 120 sentences selected from Chinese Gigaword Fifth Edition (LDC2011T13) - Simple English -- 120 sentences selected from the TIMIT Acoustic-Phonetic Continuous Speech Corpus (LDC93S1) as representative of everyday English - English Treebank -- 120 sentences selected from the WSJ subset of Treebank-3 (LDC99T42) A unique list of prompts was presented to each speaker with the following composition: - 20 calibration sentences (read by all speakers) * for the Chinese sentence lists, calibration sentences were selected to cover the maximum number of syllable types in the language - 40 shared sentences (read by 10 speakers) * for the Chinese sentence lists, shared sentences were selected to cover the maximum number of tones and tonal combinations - 80 unique sentences (read by a single speaker) 2. Collection details ===================== The corpus was recorded at Tianjin University, China in a quiet room. Fifty speakers (25 female, 25 male), all students at Tianjin University, were recruited. All speakers were born and brought up in one of the six urban districts in Tianjin, China and speak Tianjin in their everday life. Their parents and grandparents are also Tianjin locals. All 50 speakers are fluent in English, as determined by a passing score on one of the following English assessment tests: - IELTS 6.0 - TEM4 75 - CET-4 520 3. Directory Struture ===================== - data//flac/ -- FLAC files corresponding to sentences from list SENT-LIST - data//textgrids/ -- .TextGrid files containing forced alignments for sentences from list SENT-LIST - data//lab/ -- HTK label files containing word, phone, and tone-level forced alignments for sentences from list SENT-LIST - docs/recordings.tbl -- recording-level metadata; e.g., speaker identity, prompt number, sentence id, transcript - docs/sentence_types.tbl -- metadata for sentences used as prompts; e.g., target sentece text, sentence type - docs/speakers.tbl -- speaker metadata - docs/file.tbl -- listing of md5 checksums, sizes, dates, and file names for all files under "data/" - README.txt -- this file 4. File naming convention ========================= Data files (FLAC, TextGrid, HTK label files) are named according to the following convention: _ where: - SUBJECT-ID -- the anonymized id for the subject reading the sentence; e.g., "S01" - PROMPT-NUM -- the order of presentation of the sentence during the session; 1-indexed - EXT -- file extension; one of ".flac", ".TextGrid", ".words", ".phones", or ".tones" 5. File formats =============== 5.1 Audio All audio is provided in the form of 16 kHz, 16-bit mono-channel FLAC files. This was downsampled to 16 kHz from the original sample rate of 48 kHz using SoX (https://sourceforge.net/projects/sox/). 5.2 TextGrid files Each recording is accompanied by a Praat TextGrid file (https://www.fon.hum.uva.nl/praat/manual/TextGrid_file_formats.html) containing the word, phone, and tone (where applicable) segmentations resulting from forced alignment. These TextGrid files contain three interval tiers: - word -- word-level segmentation of VERBATIM transcription of subject's reading of the sentence; i.e., if the subject said something other than what was presented in the prompt, that discrepancy is recorded here - phone -- phone-level segmentation - tone -- tone-level segmentation; omitted for prompts in non-tonal languages These files may be read using Praat (https://www.fon.hum.uva.nl/praat/), ELAN (https://www.mpi.nl/corpus/html/elan/), or the Python TextGridTools package (https://github.com/hbuschme/TextGridTools). 5.3 HTK label files The segmentations contained in the TextGrid files are also distributed as HTK label files. Label files are headerless tab-delimited files containing one row per segment, each row consisting of the following 3 columns: - onset -- onset of segment in seconds from beginning of recording - offset -- offset of segment in seconds from beginning of recording - label -- segment label; e.g., a word, phone, or tone Word, phone, and tone level segmentations are each stored in separate files with the file extension indicating which tier of segmentation is present: - .words -- word-level segmentation - .phones -- phone level segmenatation - .tones -- tone level segmentation; omitted for prompts in non-tonal languages 6. Forced alignment =================== 6.1 Tianjin Forced alignment was performed using Montreal Forced Aligner (MFA): https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner All recordings were aligned using MFA v3.0 and the following acoustic model and dictionary: - acoustic model: Mandarin MFA acoustic model v3.0.0 (https://mfa-models.readthedocs.io/en/latest/acoustic/Mandarin/Mandarin%20MFA%20acoustic%20model%20v3_0_0.html#Mandarin%20MFA%20acoustic%20model%20v3_0_0) - dictionary: Mandarin (China) MFA dictionary v3.0.0 (https://mfa-models.readthedocs.io/en/latest/dictionary/Mandarin/Mandarin%20%28China%29%20MFA%20dictionary%20v3_0_0.html#Mandarin%20(China)%20MFA%20dictionary%20v3_0_0) For out-of-vocabulary (OOV) words, pronunciations were generated using a grapheme-to-phoneme (G2P) transducer trained on this dictionary. For our G2P model, we used Phonetisaurus https://github.com/AdolfVonKleist/Phonetisaurus MFA produces output in International Phonetic Alphabet (IPA) format. To make the results more accessible and consistent with standard Chinese romanization, we remapped the MFA output from IPA to Pinyin format, using the phonetic labels described in this paper: Yuan, J., Ryant, N., & Liberman, M. (2014). Automatic phonetic segmentation in Mandarin Chinese: Boundary models, glottal features and tone. In Proc. of ICASSP 2014 (pp. 2539–2543). https://doi.org/10.1109/ICASSP.2014.6854058 This mapping is performed by a left-to-right greedy algorithm that backs off to a Pinyin format pronunciation dictionary for some edge cases. In very rare cases (<0.05% of all phones), the algorithm fails and no Pinyin is prediced. When this happend, we indicate this outcome using a special "" phone (i.e., unknown); e.g.: ch e2 zh a3 Segments of time where the subject is not speaking are indicated by the "sil" word/phone. 6.2 English Forced alignment was performed using the Penn Forced Aligner: https://github.com/jaekookang/p2fa_py3 and version 0.07b of the CMU Pronouncing Dictionary (CMUdict): http://www.speech.cs.cmu.edu/cgi-bin/cmudict For out-of-vocabulary (OOV) words, pronunciations were generated using a grapheme-to-phoneme (G2P) transducer trained on this dictionary. For our G2P model, we used Phonetisaurus https://github.com/AdolfVonKleist/Phonetisaurus Segments of time where the subject is not speaking are indicated by the "sil" word/phone. 7. Metadata =========== 7.1 Recordings table (docs/recordings.tbl) Information about each recorded reading of a sentence in the corpus is stored under "docs/recordings.tbl". This is a tab-delimited file containing one row per recorded sentence reading, each row containing the following 6 columns: - Sentence List -- list sentence is from - Subject ID -- the anonymized id for the subject reading the sentence; e.g., "S01" - Prompt # -- the order of presentation of the sentence during the session; 1-indexed - Sentence ID -- an integer identifier for the sentence read by the subject; ids are unique within each sentence list - Sentence -- text presented to the subject to read - File Name -- basename of FLAC file corresponding to read sentence 7.2 Sentence types table (docs/sentence_types.tbl) The full list of sentences presented to subjects is stored under "docs/sentence_types.tbl". This is a tab-delimited file containing one row per sentence, each row containing the following 4 columns: - Sentence List -- list sentence is from - Sentence ID -- an integer identifier for the sentence; ids are unique within each sentence list - Sentence -- sentence text - Type -- sentence type; one of "Calibration", "Shared", or "Unique" 7.3 Speakers table (docs/speakers.tbl) Information about each speaker (i.e., a subject) in the corpus is stored under "docs/speakers.tbl". This is a tab-delimited file containing one row per subject, each row containing the following 4 columns: - Speaker ID -- anonymized id for the subject - Gender -- gender of the subject; either "Male" or "Female" - Age -- self-reported age in years - CET-4 -- English proficiency, as evaluated by the College English Test Band 4 (CET-4) exam Values for columns other than "Speaker ID" are based on self-reports. 8. Known issues =============== 8.1 Simple Chinese - The intended procedure for speakers in GlobalTIMIT project was to read the sentences as close to the original prompt as possible. However, the instructions provided to the speakers may have been unclear, leading them to interpret the task as requiring them to read the prompts in their natural Tianjin dialect rather than strictly adhering to the given text. Many speakers substituted words in the prompts with their Tianjin dialect equivalents. Since the corpus was transcribed verbatim, this has led to two main issues: 1. There are many variations in the final transcripts for a single prompt. For example, the Calibration sentence: 昨天 你们 队 的 篮球 比赛 结果 怎么样 resulted in 8 different variants. 2. There are significant discrepancies between the original prompts and the recorded variants. In some cases, the WER is quite large. For example, the Calibration sentence: 你 爸 这是 去 哪儿 了 had a high WER variant: 你 爸 介似 去 哪罕儿 了 - Word segmentation in verbatim transcriptions is inconsistent. - Recordings (audio and segmentation) are missing for 243 trials. 8.2 Chinese Gigaword - The intended procedure for speakers in GlobalTIMIT project was to read the sentences as close to the original prompt as possible. However, the instructions provided to the speakers may have been unclear, leading them to interpret the task as requiring them to read the prompts in their natural Tianjin dialect rather than strictly adhering to the given text. Many speakers substituted words in the prompts with their Tianjin dialect equivalents. Since the corpus was transcribed verbatim, this has led to two main issues: 1. There are many variations in the final transcripts for a single prompt. For example, the Calibration sentence: 父亲 很 奇怪 儿女 出门 为 什么 都 不 通知 一 声 resulted in 9 different variants. 2. There are discrepancies between the original prompts and the recorded variants. In some cases, the WER is quite large. For example, the Calibration sentence: 中国 银行卡 产业 近年 来 发展 迅速 had a high WER variant: 中国 银行卡 产业 最近 来 迅速 发展 - Word segmentation in verbatim transcriptions is inconsistent. 8.3 Simple English - Each unique sentence was mistakenly presented to 5 speakers (instead of 1). - Recordings (audio and segmentation) are missing for 7 trials. 9. Contacts =========== If you have questions about this data release, please contact the following LDC personnel: Neville Ryant