Diffusion-based speech generators are ubiquitous. These methods can generate very high quality synthetic speech and several recent incidents report their malicious use. To counter such misuse, synthetic speech detectors have been developed. Many of these detectors are trained on datasets which do not include diffusion-based synthesizers. In this paper, we demonstrate that existing detectors trained on one such dataset, ASVspoof2019, do not perform well in detecting synthetic speech from recent diffusion-based synthesizers. We propose the Diffusion-Based Synthetic Speech Dataset (DiffSSD), a dataset consisting of about 200 hours of labeled speech, including synthetic speech generated by 8 diffusion-based open-source and 2 commercial generators. We also examine the performance of existing synthetic speech detectors on DiffSSD in both closed-set and open-set scenarios. The results highlight the importance of this dataset in detecting synthetic speech generated from recent open-source and commercial speech generators.

Speech LLMs use speech embeddings as the prompt to a Large Language Model (LLM) and generate human readable text for the speech signal in an autoregressive manner. Teacher-forcing is a common approach used for training Speech LLMs, which is dissimilar to the procedure used during inference, creating a gap between training and inference known as exposure bias. To mitigate exposure bias, we propose Speech-N-LlaMA. Contrary to the existing Speech LLMs that have single pass through the LLM during training, Speech-N-LlaMA incorporates multi-pass training. Through multiple passes, Speech-N-LlaMA mitigates exposure bias and uses the error correction capability of LLM to improve the performance of Speech LLMs. We propose an N-pass loss and utterance level temperature sampling in Speech-N-LlaMA to achieve this. We evaluate four different model sizes on three benchmarks, and show up to 18% relative improvement in Word Error Rate (WER) compared to baseline Speech LLM, while not incurring any additional compute during inference.

Generative modeling of non-negative, discrete data, such as symbolic music, remains challenging due to two persistent limitations in existing methods. Firstly, many approaches rely on modeling continuous embeddings, which is suboptimal for inherently discrete data distributions. Secondly, most models optimize variational bounds rather than exact data likelihood, resulting in inaccurate likelihood estimates and degraded sampling quality. While recent diffusion-based models have addressed these issues separately, we tackle them jointly. In this work, we introduce the Information-Theoretic Discrete Poisson Diffusion Model (ItDPDM), inspired by photon arrival process, which combines exact likelihood estimation with fully discrete-state modeling. Central to our approach is an information-theoretic Poisson Reconstruction Loss (PRL) that has a provable exact relationship with the true data likelihood. ItDPDM achieves improved likelihood and sampling performance over prior discrete and continuous diffusion models on a variety of synthetic discrete datasets. Furthermore, on real-world datasets such as symbolic music and images, ItDPDM attains superior likelihood estimates and competitive generation quality—demonstrating a proof of concept for distribution-robust discrete generative modeling.

With recent advancements in generating synthetic speech, tools to generate high-quality synthetic speech impersonating any human speaker are easily available. Several incidents report misuse of high-quality synthetic speech for spreading misinformation and for large-scale financial frauds. Many methods have been proposed for detecting synthetic speech; however, there is limited work on localizing the synthetic segments within the speech signal. In this work, our goal is to localize the synthetic speech segments in a partially synthetic speech signal. Most existing methods for synthetic speech localization obtain features from either the time domain waveform or the spectrogram representation of the speech signal. In this work, we propose Multi-Domain ResNet Transformer (MDRT) that obtains multi-domain features from both the time domain and the spectrogram representation of a speech signal to localize synthetic speech segments. MDRT uses transformer neural networks to obtain multi-domain features and processes them using a ResNet-style neural network. We use the PartialSpoof dataset to examine the performance of MDRT on localizing synthetic speech segments of varying duration. Our results show that MDRT performs better than several existing synthetic speech localization methods.

Deep learning methods can generate high-quality synthetic speech which is perceptually indistinguishable from real human speech. Synthetic speech can be maliciously used for fraud. Synthetic speech detection methods have been proposed which perform well on ASVspoof2019 and ASVspoof2021 Datasets. These datasets consist of synthetic speech from conventional neural network speech generators. Recently, many voice cloning methods have been proposed which use diffusion models and generative adversarial networks for high-quality speech synthesis. In this work, we present a new synthetic speech dataset containing 25,000 synthetic speech signals for 11 distinct speakers, with a total duration of 52 hours. We have developed this dataset using 5 recent diffusion model-based synthetic speech generators. These generators can clone a speaker's voice from text using only a few minutes of their real speech. We evaluate 6 of the best synthetic speech detectors that work well on the ASVspoof2019 Dataset on this new dataset, and demonstrate their performance using Equal Error Rate (EER).

Synthetic human speech signals have become very easy to generate given modern text-to-speech methods. When these signals are shared on social media they are often compressed using the Advanced Audio Coding (AAC) standard. Our goal is to study if a small set of coding metadata contained in the AAC compressed bit stream is sufficient to detect synthetic speech. This would avoid decompressing of the speech signals before analysis. We call our proposed method AAC Synthetic Speech Detection (ASSD). ASSD extracts information from the AAC compressed bit stream without decompressing the speech signal. ASSD analyzes the information using a transformer neural network. In our experiments, we compressed the ASVspoof2019 dataset according to the AAC standard using different data rates. We compared the performance of ASSD to a time domain based and a spectrogram based synthetic speech detection methods. We evaluated ASSD on approximately 71k compressed speech signals. The results show that our proposed method typically only requires 1000 bits per speech block/frame from the AAC compressed bit stream to detect synthetic speech. This is much lower than other reported methods. Our method also had a 9.7 percentage points higher detection accuracy compared to existing methods.

Tools to generate high quality synthetic speech that is perceptually indistinguishable from speech recorded from hu-man speakers are easily available. Many incidents report misuse of synthetic speech for spreading misinformation and committing financial fraud. Several approaches have been proposed for detecting synthetic speech. Many of these approaches use deep learning methods without providing reasoning for the decisions they make. This limits the explainability of these approaches. In this paper, we use disentangled representation learning for developing a synthetic speech detector. We propose Disentangled Spectrogram Variational Auto Encoder (DSVAE) which is a two stage trained variational autoencoder that processes spectrograms of speech to generate features that disentangle synthetic and bona fide speech. We evaluated DSVAE using the ASVspoof2019 dataset. Our experimental results show high accuracy (> 98%) on detecting synthetic speech from 6 known and 10 unknown speech synthesizers. Further, the visualization of disentangled features obtained from DSVAE provides rea-soning behind the working principle of DSVAE, improving its explainability. DSVAE performs well compared to several existing methods. Additionally, DSVAE works in practical scenarios such as detecting synthetic speech uploaded on social platforms and against simple attacks such as removing silence regions.

Many deep learning synthetic speech generation tools are readily available. The use of synthetic speech has caused financial fraud, impersonation of people, and misinformation to spread. For this reason forensic methods that can detect synthetic speech have been proposed. Existing methods often overfit on one dataset and their performance reduces substantially in practical scenarios such as detecting synthetic speech shared on social platforms. In this paper we propose, Patched Spectrogram Synthetic Speech Detection Transformer (PS3DT), a synthetic speech detector that converts a time domain speech signal to a mel-spectrogram and processes it in patches using a trans-former neural network. We evaluate the detection performance of PS3DT on ASVspoof2019 dataset. Our experiments show that PS3DT performs well on ASVspoof2019 dataset compared to other approaches using spectrogram for synthetic speech detection. We also investigate generalization performance of PS3DT on In-the-Wild dataset. PS3DT generalizes well than several existing methods on detecting synthetic speech from an out-of-distribution dataset. We also evaluate robustness of PS3DT to detect telephone quality synthetic speech and synthetic speech shared on social platforms (compressed speech). PS3DT is robust to compression and can detect telephone quality synthetic speech better than several existing methods.

The malicious use of synthetic speech has increased with the recent availability of speech generation tools. It is important to determine whether a speech signal is authentic (spoken by a human) or is synthesized and to determine the generation method used to create it. Identifying the synthesis method is known as synthetic speech attribution. In this paper, we propose the use of a transformer deep learning method that analyzes mel-spectrograms for synthetic speech attribution. Our method known as Patchout Spectrogram Attribution Transformer (PSAT) can distinguish new, unseen speech generation methods from those seen during training. PSAT demonstrates high performance in attributing synthetic speech signals. Evaluation on the DARPA SemaFor Audio Attribution Dataset and the ASVSpoof2019 Dataset shows that our method achieves more than 95% accuracy in synthetic speech attribution and performs better than existing deep learning approaches.

Many speech signals are compressed with MP3 to reduce the data rate. In many synthetic speech detection methods the spectrogram of the speech signal is used. This usually requires the speech signal to be fully decompressed. We show that the design of MP3 compression allows one to approximate the spectrogram of the MP3 compressed speech efficiently without fully decoding the compressed speech. We denote the spectograms obtained using our proposed approach by Efficient Spectrograms (E-Specs). E-Spec can reduce the complexity of spectrogram computation by ~77.60 percentage points (p.p.) and save ~37.87 p.p. of MP3 decoding time. E-Spec bypasses the reconstruction artifacts introduced by the MP3 synthesis filterbank, which makes it useful in speech forensics tasks. We tested E-Spec in the synthetic speech detection, where a detector is asked to determine whether a speech signal is synthesized or recorded from a human. We examined 4 different neural network architectures to evaluate the performance of E-Spec compared to speech features extracted from the fully decoded speech signal. E-Spec achieved the best synthetic speech detection performance for 3 architectures; it also achieved the best overall detection performance across architectures. The computation of E-Spec is an approximation to Short Time Fourier Transform (STFT). E-Spec can be extended to other audio compression methods.