A test aimed at homophones and wordplay
The evaluation reports two gains against baseline: recall of intended HumourPhone substitutions was more than 5% higher, and target-span character error rate for homophone correction was 4.35% lower. The evaluation was inference-only, with no additional training or fine-tuning.
HumourPhone in this study means intentional homophone wordplay. The proposed framework is built to detect Mandarin homophone spans in speech and recover either an ordinary homophone correction or an intentional HumourPhone substitution. Its adapter combines a part-of-speech-aware homophone detector, a branch-based large-language-model prompt builder, candidate generation and a BERT-based semantic output selector.
Two datasets, two kinds of challenge
The main CN Dataset was formed from AISHELL-3 and CommonVoice 25 Chinese. Invalid targets were removed before the remaining targets were grouped with pinyin-based homophone categories. It contains 1,303 homophone groups, 85,706 target occurrences and 125.73 hours of speech. The separate HumourPhone set contains 80 groups, 80 occurrences and 0.08 hours.
HumourPhone examples came from Chinese internet sources. Image text was manually extracted and normalized, and Mandarin audio was synthesized with Microsoft Edge text-to-speech. The evaluation compared standard automatic speech recognition, prompt-only decoding and the proposed adapter without additional training or fine-tuning.
Performance was measured with target-span character error rate, or T-CER. It is based on character-level Levenshtein distance, which counts the character edits between an output and the target. The evaluation also used Recovery Rate, or RR, for the share of initially incorrect targets that were corrected; Detection Rate, or DetR, based on overlap between predicted and target spans; Mean Uncommon Recall, or MUR, based on recall; and utterance-level and target-span semantic similarity, known as USS and TSS.
Results shifted with the model and prompt
On HumourPhone, a Chinese-guidelines prompt labeled prompt 3 had an overall T-CER of 13.88%, compared with 15.21% for zero-shot Qwen3-ASR. Its near-homophone T-CER was 9.69% versus 11.61%, and its full-homophone T-CER was 15.92% versus 16.96%. Near-homophone target-span semantic similarity was 0.9823 versus 0.9793, while overall Mean Uncommon Recall was 0.4547 versus 0.4043.
Other prompts produced different results. Prompt 1, an English-basic prompt, had an overall T-CER of 15.97%, while prompt 4, a Chinese-guidelines prompt with an example, had 14.04%. Both were above prompt 3's 13.88% in the same HumourPhone comparison.
On the CN Dataset, the adapter-condition results for Whisper-v3 were below baseline on both frequency groups. T-CER was 5.07% versus 6.43% for common forms and 20.39% versus 24.74% for uncommon forms. Detection Rate was 37.22% for common forms and 47.27% for uncommon forms, while Recovery Rate was 7.18% and 20.87%, respectively.
For Qwen3-ASR, the adapter-condition T-CER was 12.85% versus 13.56% baseline on uncommon forms, but 3.42% versus 2.19% on common forms. Detection Rate was 44.42% for common forms and 50.67% for uncommon forms. Recovery Rate was 2.79% and 8.94%, respectively.
An emotion detector used in the evaluation was accurate on 97.0% of raw Qwen3-ASR transcripts and 76.0% of raw Whisper-Large-v3 transcripts.
Full homophones remain the warning sign
The study's clearest caution concerns full homophones. The authors state that gains remained limited, especially for full homophones, and that prompting alone was insufficient for robust HumourPhone recovery.
That caution sits alongside a small, synthetic wordplay test. HumourPhone contains 80 occurrences across 80 groups and 0.08 hours of audio. Its examples were collected from Chinese internet sources, with image text manually extracted and normalized before Mandarin audio was synthesized with Microsoft Edge text-to-speech.
Taken together, the comparisons are condition-specific. Prompt 3 was below zero-shot Qwen3-ASR on overall HumourPhone T-CER, while the CN Dataset comparisons showed lower uncommon-form T-CER for both Whisper-v3 and Qwen3-ASR but higher common-form T-CER for Qwen3-ASR under the adapter. The authors nevertheless describe the gains as limited, particularly for full homophones.
The work is an arXiv preprint identified as version 2 and dated 27 August 2026. It is an inference-only evaluation, and the authors say prompting alone was insufficient for robust HumourPhone recovery.
Paper data and sources
Original title: Mandarin Humorous Homophone Recognition and Disambiguation in Automatic Speech Recognition
Authors: Sicheng Jin, Jinghao Chen, Mostafa Shahin et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text