On the use of I-vectors and average voice model for voice conversion without parallel data

摘要

Recently, deep and/or recurrent neural networks (DNNs/RNNs) have been employed for voice conversion, and have significantly improved the performance of converted speech. However, DNNs/RNNs generally require a large amount of parallel training data (e.g., hundreds of utterances) from source and target speakers. It is expensive to collect such a large amount of data, and impossible in some applications, such as cross-lingual conversion. To solve this problem, we propose to use average voice model and i-vectors for long short-term memory (LSTM) based voice conversion, which does not require parallel data from source and target speakers. The average voice model is trained using other speakers' data, and the i-vectors, a compact vector representing the identities of source and target speakers, are extracted independently. Subjective evaluation has confirmed the effectiveness of the proposed approach.