Kaldi Phoneme Recognition Method Validated for Verifying TTS Pronunciation of Polyphonic Characters
A Kaldi-based phoneme recognition method for verifying TTS pronunciation of polyphonic and rare Chinese characters was validated as feasible. The method feeds TTS audio into Kaldi speech recognition and compares the decoded phoneme sequence, rather than recognized Chinese characters, against the target character's correct pronunciation. A reproduction tutorial covers CPU-only Kaldi installation, the Vosk Chinese acoustic model, and a test case checking whether 攒 in 攒竹穴 is read as cuán rather than zǎn.
Event version 1.
Key insights
Phoneme decoding avoids word-level recognition bias
Word-level recognition is influenced by the language model, so a mispronounced word may still be output as a contextually plausible word. Phonemes come from the acoustic decoding path, where pronunciation variants compete acoustically, so the selected phoneme sequence reflects the actual sound produced.
CPU-only Kaldi setup via conda-forge
The tutorial installs Kaldi 5.5.1172 CPU build from conda-forge rather than compiling from source. It requires Linux x86_64, at least 4 GB memory, and about 4 GB disk for the model and environment.
Vosk large Chinese model required
The method uses the Vosk vosk-model-cn-0.22 large Chinese model, a packaged Kaldi chain model with acoustic model, HCLG decoding graph, word list, and i-vector extractor. The small model's private lookahead graph format causes segfaults in upstream Kaldi decoders.
Calibration sentence anchors phoneme IDs
A calibration sentence made of unambiguous words is used to anchor phoneme numbering, covering the target syllable's initial, easily confused initials, tones, and same-structure syllables. For cuán, words like 全/团, 咱, 敢, 在/做, and 村/侧 provide anchors.
Test case checks 攒竹穴 pronunciation
The test sentence places 攒竹穴 in natural context with the target character appearing twice for cross-checking. The correct reading is cuán, not zǎn.