Abstract
This paper presents a method of sequence-to-sequence (seq2seq) voice conversion using non-parallel training data. In this method, disentangled linguistic and speaker representations are extracted from acoustic features, and voice conversion is achieved by preserving the linguistic representations of source utterances while replacing the speaker representations with the target ones. Our model is built under the framework of encoder-decoder neural networks. A recognition encoder is designed to learn the disentangled linguistic representations with two strategies. First, phoneme transcriptions of training data are introduced to provide the references for leaning linguistic representations of audio signals. Second, an adversarial training strategy is employed to further wipe out speaker information from the linguistic representations. Meanwhile, speaker representations are extracted from audio signals by a speaker encoder. The model parameters are estimated by two-stage training, including a pretraining stage using a multi-speaker dataset and a fine-tuning stage using the dataset of a specific conversion pair. Since both the recognition encoder and the decoder for recovering acoustic features are seq2seq neural networks, there are no constraints of frame alignment and frame-by-frame conversion in our proposed method. Experimental results showed that our method obtained higher similarity and naturalness than the best non-parallel voice conversion method in Voice Conversion Challenge 2018. Besides, the performance of our proposed method was closed to the state-of-the-art parallel seq2seq voice conversion method.
Introduction
Voice Conversion (VC) is the task of changing the speaker identity of an utterance while keeping the underlying message intact. Traditionally, the most successful models required parallel data—recordings of different people saying the exact same sentences. However, such datasets are rare and difficult to collect in real-world scenarios.
Non-parallel VC is more practical but significantly harder due to the challenge of "disentanglement": separating linguistic content from speaker identity. Furthermore, most existing non-parallel methods operate frame-by-frame, limiting their ability to adjust duration or prosody. This paper bridges this gap by bringing the flexibility of seq2seq modeling to the non-parallel VC domain.
Methods
The authors propose an encoder-decoder architecture designed to isolate speech components into distinct linguistic and speaker spaces.
Disentangled Representations
- Recognition Encoder ($E^r$): This component takes acoustic features and compresses them into a linguistic representation. It uses a "pyramid" BLSTM to reduce time resolution, effectively mapping high-frequency audio frames to lower-frequency phoneme-like units.
- Text Encoder ($E^t$): During training, phoneme transcriptions are fed into this encoder. A contrastive loss forces the output of the Recognition Encoder to match the output of the Text Encoder, ensuring they share the same linguistic space.
- Speaker Encoder ($E^s$): A separate network extracts a fixed embedding representing the target speaker's unique vocal characteristics.
Adversarial Training and Decoding
To ensure the linguistic representation is truly "speaker-blind," an auxiliary classifier ($C^s$) attempts to guess the speaker's identity from it. The Recognition Encoder is trained to "fool" this classifier, scrubbing away residual speaker-specific information. Finally, the Decoder ($D^a$) takes the cleaned linguistic embedding and the target speaker embedding to reconstruct the Mel-spectrogram. Because it is a seq2seq model with an attention mechanism, it can predict durations independently of the source audio.
Key Results
Experimental results demonstrated that the proposed method achieved higher similarity and naturalness than the top-performing non-parallel method from the Voice Conversion Challenge 2018 (CycleGAN-VC).
- Naturalness and Similarity: In subjective listening tests, the performance was nearly indistinguishable from state-of-the-art models using parallel data.
- Objective Metrics: The model achieved a Mel-cepstrum distortion (MCD) of ~3.0 dB, outperforming other non-parallel baselines.
- Data Efficiency: The model proved robust even with limited data, outperforming the VCC2018 baseline using only 100 non-parallel utterances compared to the baseline's 500.
Conclusion
This work successfully demonstrates that combining seq2seq frameworks with adversarial training and phoneme-level supervision can produce high-quality non-parallel voice conversion. By effectively disentangling linguistic content from speaker identity, the model allows for flexible duration modification and natural-sounding speech. Future work will focus on "one-shot" conversion and better capturing complex speaking styles and emotional nuances.
Summary updated