PTC-Bias
A context biasing framework based on phoneme-level temporal competition. It employs a two-stage design—retrieval followed by correction—to improve rare-word recognition accuracy in speech large language models without requiring additional SpeechLLM forward computation passes.
Timeline (last 90 days)
PTC-Bias is a two-stage context biasing framework based on phoneme-level temporal competition, designed to improve the recognition of rare words in speech large language models.
PTC-Bias splits the process into two stages: PTC retrieval and PTC correction.
PTC retrieval performs frame-synchronous phoneme decoding during the prefill stage, allowing candidate word pronunciations to compete on the time axis, filtering out a compact shortlist of candidate bias words and locating the corresponding speech intervals.
PTC correction performs local competition within speech intervals for segments where retrieved candidate words do not match the transcription text, specifically targeting near-homophone words and segmentation errors.
The two stages of PTC-Bias share the same set of phoneme posterior probabilities and do not require additional SpeechLLM forward passes.
The paper validates on the LibriSpeech dataset, covering two different SpeechLLMs, with the maximum bias word list size reaching 2,000 words.
When using the Prompt-SLAM-ASR-7B model with a 2,000-word bias list, B-WER on the test-clean set was reduced by 23.4% relative to the CTC-Filter baseline.
On the test-other set, B-WER was reduced by 23.9% relative to the CTC-Filter baseline.
While PTC-Bias reduces B-WER, U-WER (unbiased word error rate) remains largely unchanged.
PTC-Bias current experiments are primarily based on the English LibriSpeech dataset and specific SpeechLLMs; its performance in other languages such as Chinese, real noisy environments, and larger bias word lists remains to be validated.
All Facts (10)
PTC-Bias is a two-stage context biasing framework based on phoneme-level temporal competition, designed to improve the recognition of rare words in speech large language models.
50%UnverifiedPTC-Bias splits the process into two stages: PTC retrieval and PTC correction.
50%UnverifiedPTC retrieval performs frame-synchronous phoneme decoding during the prefill stage, allowing candidate word pronunciations to compete on the time axis, filtering out a compact shortlist of candidate bias words and locating the corresponding speech intervals.
50%UnverifiedPTC correction performs local competition within speech intervals for segments where retrieved candidate words do not match the transcription text, specifically targeting near-homophone words and segmentation errors.
50%UnverifiedThe two stages of PTC-Bias share the same set of phoneme posterior probabilities and do not require additional SpeechLLM forward passes.
50%UnverifiedThe paper validates on the LibriSpeech dataset, covering two different SpeechLLMs, with the maximum bias word list size reaching 2,000 words.
50%UnverifiedWhen using the Prompt-SLAM-ASR-7B model with a 2,000-word bias list, B-WER on the test-clean set was reduced by 23.4% relative to the CTC-Filter baseline.
50%UnverifiedOn the test-other set, B-WER was reduced by 23.9% relative to the CTC-Filter baseline.
50%UnverifiedWhile PTC-Bias reduces B-WER, U-WER (unbiased word error rate) remains largely unchanged.
50%UnverifiedPTC-Bias current experiments are primarily based on the English LibriSpeech dataset and specific SpeechLLMs; its performance in other languages such as Chinese, real noisy environments, and larger bias word lists remains to be validated.
50%