Characters in the dataset are shown with their reconstructed Old Chinese IPA and English gloss. Brackets (e.g. [t]) indicate uncertainty in the reconstruction. Characters not in the Baxter–Sagart dataset appear as ? with a brief pause in the audio. Only the first reconstruction is played when a character has multiple readings.
Old Chinese is the oldest stage of the Chinese language for which we have substantial written evidence, roughly 1250–256 BCE — spanning the late Shang dynasty through the Warring States period. It is the language of the Shijing (Classic of Poetry), the Shujing, and the earliest philosophical texts. It predates Middle Chinese (the language of the Tang dynasty poets) by roughly a millennium, and differs from modern Mandarin as much as Latin differs from French.
Because Old Chinese was never written with a phonetic script, its pronunciation cannot be directly observed. Linguists reconstruct it by working backwards from multiple lines of evidence: the rhyme patterns of the Shijing, the phonetic components of Chinese characters, loanwords between Old Chinese and neighbouring languages, the sound changes attested in later Middle Chinese, and data from modern dialects — especially the Min languages, which preserve archaic features lost elsewhere.
The Baxter–Sagart system (William Baxter & Laurent Sagart, Old Chinese: A New Reconstruction, Oxford University Press 2014) is the current standard in Western sinological research. It reconstructs roughly 5,000 morphemes and introduces features such as pharyngealized initials and uvular stops not found in earlier systems. Reconstructed forms are written in IPA and conventionally prefixed with an asterisk, e.g. *truk for 竹 (bamboo).
The IPA strings from the Baxter–Sagart dataset are passed directly to Azure Neural TTS via SSML <phoneme alphabet="ipa"> tags, bypassing any text-to-phoneme step. This is architecturally correct for a dead language: no text-to-phoneme model exists for Old Chinese, so phoneme-level input is the only approach that preserves fidelity to the source data. Audio is pre-synthesised for all characters in the dataset. Features with no equivalent in the voice's phoneme inventory — such as pharyngealization (ˤ) — are silently dropped. Treat the audio as an impression, not a faithful reconstruction.
Pronunciation data is sourced from Wiktionary's Baxter–Sagart Lua modules (CC BY-SA), covering ~4,033 characters. The authoritative dataset is available directly from the authors' site.