Abstract
This paper introduces "Speak Like a Dog," a novel voice conversion (VC) task that transforms human speech into dog-like vocalizations while attempting to preserve the original linguistic content. By utilizing non-parallel VC frameworks like StarGAN-VC and ACVAE-VC, the researchers found that using mel-spectrograms and smaller CNN kernel sizes significantly improves "dog-likeness." However, maintaining high intelligibility remains a major technical challenge, highlighting a fundamental trade-off in human-to-non-human voice conversion.
Introduction
Traditional Voice Conversion (VC) focuses on mapping one human voice to another. However, there is a growing interest in Human-to-Non-Human Voice Conversion (H2NH-VC) for applications in cinema, gaming, and creative arts—imagine a monster or a pet in a movie "speaking" lines while retaining their animalistic character.
The primary hurdle in H2NH-VC is the lack of parallel data. In standard VC, models are often trained on pairs of speakers saying the same sentences. Since "there is no dog in this world that speaks human languages," researchers must rely on non-parallel VC methods. This study investigates whether existing state-of-the-art non-parallel models can bridge the gap between human phonemes and canine barks.
Methods
The authors evaluated two representative non-parallel VC architectures:
- StarGAN-VC: A Generative Adversarial Network (GAN) based approach that allows for many-to-many mappings.
- ACVAE-VC: An Auxiliary Classifier Variational Autoencoder-based method known for stable training and fast convergence.
The research explored two types of acoustic features:
- Mel-cepstral coefficients (MCC): Traditionally used in human VC because they represent the human vocal tract shape.
- Mel-spectrograms: These include more harmonic structure and raw spectral information.
Additionally, they conducted a sensitivity analysis on CNN kernel sizes within the discriminator and domain classifier. They hypothesized that because dog barks are shorter and more transient than human speech, the temporal receptive field of the network (controlled by kernel size) would be a critical factor in capturing "dog-ness."
Datasets
- Human Speech: The ATR digital sound database, featuring 503 utterances each from four professional announcers (two male, two female).
- Dog Voices: A newly constructed dataset compiled from Freesound and YouTube. It contains 792 adult dog sounds and 288 puppy sounds, filtered to remove noise and extreme volume fluctuations.
Key Insights and Ideas
- Mel-spectrograms are superior for H2NH-VC: While MCCs are standard for human-to-human tasks, they failed to capture the "dog-like" quality. Mel-spectrograms, which preserve more raw spectral detail, were significantly better at representing the non-harmonic, noisy characteristics of a bark.
- Smaller Kernels for Shorter Sounds: Contrary to many deep learning trends where "bigger is better," increasing the CNN kernel size actually decreased performance. Smaller kernels were more effective at capturing the rapid, transient nature of canine vocalizations.
- The Intelligibility Trade-off: There is a fundamental tension between "dog-likeness" and "clarity." The features that make a sound recognizable as a dog (noise, specific spectral envelopes) often mask the subtle acoustic cues required for human speech recognition.
Key Results
The researchers used Mean Opinion Scores (MOS) and Character Error Rate (CER) to evaluate the results:
- Dog-likeness: Both StarGAN-VC and ACVAE-VC using mel-spectrograms achieved high scores (approx. 4.2/5.0), whereas MCC-based versions struggled (approx. 1.2–2.0).
- Clarity and Intelligibility: This was the weak point. StarGAN-VC with mel-spectrograms achieved the highest clarity among converted voices, but only scored 2.04/5.0.
- Transcription Accuracy: The CER was very high (above 0.80 for most dog-like configurations), meaning listeners could rarely transcribe the converted speech accurately. ACVAE-VC with MCC had the lowest error but sounded the least like a dog.
Limitations and Future Work
The authors acknowledge that while they successfully "dog-ified" the audio, the linguistic information was largely lost. The current models struggle to balance the source's linguistic structure with the target's non-human acoustic texture.
Future work suggests exploring specialized vocoders (the components that turn acoustic features back into waveforms) that are not biased toward human speech production models. Additionally, finding a better way to disentangle linguistic content from speaker identity in the latent space remains a priority for making "speaking animals" a reality in digital media.
Summary updated