In automated, inference-only tests, a task-conditioned adapter for Mandarin speech recognition showed gains on ordinary homophone correction and on a synthetic task built around humorous wordplay. The task concerns cases where the same spoken sound can map to different written forms. The abstract reports that HumourPhone recall was more than 5% higher than baseline, while target-span character error rate, or T-CER, was 4.35% lower for homophone correction.
A system built to choose the intended form
The study asks whether an ASR adaptation framework can identify intentional Mandarin homophone wordplay, restore the intended written form and correct ordinary homophones at the same time. Its adapter detects suspicious spans, routes utterances to homophone or HumourPhone branches, generates candidate transcriptions and selects a final output using contextual information and a MacBERT-based naturalness score.
The evaluation compared Qwen3-ASR and Whisper-v3 outputs with a prompt-only Qwen3-LLM baseline and the proposed adapter. In the reported setup, Qwen2.5-7B-Instruct generated two alternative transcriptions alongside the original transcription under deterministic decoding, with the decoding parameter set to zero and a maximum of 256 new tokens.
The study tracked several measures, including target-span character error rate, detection rate, recovery rate, Mean Uncommon Recall and semantic-similarity scores at both utterance and target-span level.
The test sets were not equal in size
The merged CN Dataset supplied the larger test bed: 1,303 homophone groups and 85,706 occurrences across 125.73 hours. The HumourPhone set was much smaller, with 80 groups and 80 occurrences over 0.08 hours.
HumourPhone was constructed from internet text and images converted into 80 Microsoft Edge TTS audio samples. The paper describes it as a controlled evaluation set, not a comprehensive natural-speech corpus.
The gains were uneven
Against Whisper-v3, the adapter comparison lowered T-CER from 6.43% to 5.07% for common forms and from 24.74% to 20.39% for uncommon forms. The larger reported change was in the uncommon group. Detection and recovery rates were 37.22% and 7.18% for common forms, compared with 47.27% and 20.87% for uncommon forms.
Qwen results were more mixed. Prompt-only Qwen3-LLM decoding nudged overall T-CER up from 3.60% to 3.70%. With the adapter, Qwen3-ASR T-CER fell from 13.56% to 12.85% on uncommon forms, but rose from 2.19% to 3.42% on common forms. The corresponding detection and recovery rates were 44.42% and 2.79% for common forms, and 50.67% and 8.94% for uncommon forms.
On the HumourPhone task, prompt 3 was the strongest listed variant. Overall mean T-CER fell from 15.21% to 13.88%, a reported relative reduction of 8.7%; for near-homophones, it fell from 11.61% to 9.69%, a 16.5% relative reduction. Near-homophone target-span semantic similarity rose from 0.9793 to 0.9823, while overall Mean Uncommon Recall increased from 0.4043 to 0.4547.
The caveats matter
The authors interpret the pattern as showing that standard ASR favors common homophone forms, simple prompting offers limited gains and targeted adapter-based post-processing is promising for ordinary correction and HumourPhone recovery, particularly on uncommon forms.
The paper also identifies important limits: the HumourPhone data are small and synthetic, no human evaluation was reported, LLM inference is costly and the system sometimes overcorrects. It is a preprint, identified in the supplied record as arXiv:2608.25384v4 in eess.AS, dated 9 Sep 2026.
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