Re-scoring speech recognition method, device, equipment and medium

• Main Author: BAN ZHIHUA
• Format: Patent
• Language: Chinese; English
• Published: 10.05.2024
• Subjects: ACOUSTICS; MUSICAL INSTRUMENTS; PHYSICS; SPEECH ANALYSIS OR SYNTHESIS; SPEECH OR AUDIO CODING OR DECODING; SPEECH OR VOICE PROCESSING; SPEECH RECOGNITION

The invention relates to the technical field of voice recognition, in particular to a re-scoring voice recognition method and device, equipment and a medium. The method comprises the following steps: acquiring target voice data; inputting the target voice data into a trained re-scoring voice recognition model for re-scoring to obtain re-scores of a plurality of candidate texts corresponding to the target voice data; screening out the candidate text with the highest matching degree from the plurality of candidate texts as an identification result of the target voice data; wherein the re-scoring speech recognition model comprises an encoder, a compression network, a CTC decoder and an attention decoder. The method is simple and efficient, delay can be reduced, rapid and accurate re-scoring is realized, and the speech recognition efficiency is improved. 本申请涉及语音识别技术领域，更为具体来说，本申请涉及一种重打分语音识别方法、装置、设备及介质。所述方法包括：获取目标语音数据；将所述目标语音数据输入训练好的重打分语音识别模型进行重打分，得到所述目标语音数据对应的多个候选文本的重打分；从所述多个候选文本中筛选出匹配度最高的候选文本作为目标语音数据的识别结果；其中，所述重打