We propose a method for the task of text-conditioned speech insertion, i.e. inserting a speech sample in an input speech sample, conditioned on the corresponding complete text transcript. An example use case of the task would be to update the speech audio when corrections are done on the corresponding text transcript. The proposed method follows a transformer-based non-autoregressive approach that allows speech insertions of variable lengths, which are dynamically determined during inference, based on the text transcript and tempo of the available partial input. It is capable of maintaining the speaker's voice characteristics, prosody and other spectral properties of the available speech input. Results from our experiments and user study on LibriTTS show that our method outperforms baselines based on an existing adaptive text to speech method. We also provide numerous qualitative results to appreciate the quality of the output from the proposed method.

Automatic detection of prominence at the word and syllable-levels is critical for building computer-assisted language learning systems. It has been shown that prosody embeddings learned by the current state-of-the-art (SOTA) text-to-speech (TTS) systems could generate word- and syllable-level prominence in the synthesized speech as natural as in native speech. To understand the effectiveness of prosody embeddings from TTS for prominence detection under nonnative context, a comparative analysis is conducted on the embeddings extracted from native and non-native speech considering the prominence-related embeddings: duration, energy, and pitch from a SOTA TTS named FastSpeech2. These embeddings are extracted under two conditions considering: 1) only text, 2) both speech and text. For the first condition, the embeddings are extracted directly from the TTS inference mode, whereas for the second condition, we propose to extract from the TTS under training mode. Experiments are conducted on native speech corpus: Tatoeba, and non-native speech corpus: ISLE. For experimentation, word-level prominence locations are manually annotated for both corpora. The highest relative improvement on word \& syllable-level prominence detection accuracies with the TTS embeddings are found to be 13.7% & 5.9% and 16.2% & 6.9% compared to those with the heuristic-based features and self-supervised Wav2Vec-2.0 representations, respectively.

Singing voice synthesis (SVS) aims to produce high-fidelity singing audio from music scores, requiring a detailed understanding of notes, pitch, and duration, unlike text-to-speech tasks. Although diffusion models have shown exceptional performance in various generative tasks like image and video creation, their application in SVS is hindered by time complexity and the challenge of capturing acoustic features, particularly during pitch transitions. Some networks learn from the prior distribution and use the compressed latent state as a better start in the diffusion model, but the denoising step doesn't consistently improve quality over the entire duration. We introduce RDSinger, a reference-based denoising diffusion network that generates high-quality audio for SVS tasks. Our approach is inspired by Animate Anyone, a diffusion image network that maintains intricate appearance features from reference images. RDSinger utilizes FastSpeech2 mel-spectrogram as a reference to mitigate denoising step artifacts. Additionally, existing models could be influenced by misleading information on the compressed latent state during pitch transitions. We address this issue by applying Gaussian blur on partial reference mel-spectrogram and adjusting loss weights in these regions. Extensive ablation studies demonstrate the efficiency of our method. Evaluations on OpenCpop, a Chinese singing dataset, show that RDSinger outperforms current state-of-the-art SVS methods in performance.

An increasing number of Chinese people are troubled by different degrees of visual impairment, which has made the modal conversion between a single image or video frame in the visual field and the audio expressing the same information a research hotspot. Deep learning technologies such as OCR+Vocoder and Im2Wav enable English audio synthesis or image-to-sound matching in a self-supervised manner. However, the audio data used for training is limited and English is not universal for visually impaired people with different educational levels. Therefore, for the sake of solving the problems of data volume and language applicability to improve the reading efficiency of visually impaired people, a set of image-to-speech framework CLIP-KNN-Fastspeech2 based on the Chinese context was constructed. The framework integrates multiple basic models and adopts the strategy of independent pre-training and joint fine-tuning. First, the Chinese CLIP and Fastspeech2 text-to-speech models were pre-trained on two public datasets, MUGE and Baker, respectively, and their convergence was verified. Subsequently, joint fine-tuning was performed using a self-built Braille image dataset. Experimental results on multiple public datasets such as VGGSound, Flickr8k, ImageHear, and the self-built Braille dataset BIT-DP show that the model has improved objective indicators such as BLEU4,FAD(Fr\'echet Audio Distance), WER(Word Error Ratio), and even inference speed. This verifies that the constructed model still has the ability to synthesize high-quality speech under limited data, and also proves the effectiveness of the joint training strategy that integrates multiple basic models.

Acoustic features play an important role in improving the quality of the synthesised speech. Currently, the Mel spectrogram is a widely employed acoustic feature in most acoustic models. However, due to the fine-grained loss caused by its Fourier transform process, the clarity of speech synthesised by Mel spectrogram is compromised in mutant signals. In order to obtain a more detailed Mel spectrogram, we propose a Mel spectrogram enhancement paradigm based on the continuous wavelet transform (CWT). This paradigm introduces an additional task: a more detailed wavelet spectrogram, which like the post-processing network takes as input the Mel spectrogram output by the decoder. We choose Tacotron2 and Fastspeech2 for experimental validation in order to test autoregressive (AR) and non-autoregressive (NAR) speech systems, respectively. The experimental results demonstrate that the speech synthesised using the model with the Mel spectrogram enhancement paradigm exhibits higher MOS, with an improvement of 0.14 and 0.09 compared to the baseline model, respectively. These findings provide some validation for the universality of the enhancement paradigm, as they demonstrate the success of the paradigm in different architectures.

We present ParrotTTS, a modularized text-to-speech synthesis model leveraging disentangled self-supervised speech representations. It can train a multi-speaker variant effectively using transcripts from a single speaker. ParrotTTS adapts to a new language in low resource setup and generalizes to languages not seen while training the self-supervised backbone. Moreover, without training on bilingual or parallel examples, ParrotTTS can transfer voices across languages while preserving the speaker specific characteristics, e.g., synthesizing fluent Hindi speech using a French speaker's voice and accent. We present extensive results in monolingual and multi-lingual scenarios. ParrotTTS outperforms state-of-the-art multi-lingual TTS models using only a fraction of paired data as latter.

While FastSpeech2 aims to integrate aspects of speech such as pitch, energy, and duration as conditional inputs, it still leaves scope for richer representations. As a part of this work, we leverage representations from various Self-Supervised Learning (SSL) models to enhance the quality of the synthesized speech. In particular, we pass the FastSpeech2 encoder's length-regulated outputs through a series of encoder layers with the objective of reconstructing the SSL representations. In the SALTTS-parallel implementation, the representations from this second encoder are used for an auxiliary reconstruction loss with the SSL features. The SALTTS-cascade implementation, however, passes these representations through the decoder in addition to having the reconstruction loss. The richness of speech characteristics from the SSL features reflects in the output speech quality, with the objective and subjective evaluation measures of the proposed approach outperforming the baseline FastSpeech2.

State-of-the-art speech synthesis models try to get as close as possible to the human voice. Hence, modelling emotions is an essential part of Text-To-Speech (TTS) research. In our work, we selected FastSpeech2 as the starting point and proposed a series of modifications for synthesizing emotional speech. According to automatic and human evaluation, our model, EmoSpeech, surpasses existing models regarding both MOS score and emotion recognition accuracy in generated speech. We provided a detailed ablation study for every extension to FastSpeech2 architecture that forms EmoSpeech. The uneven distribution of emotions in the text is crucial for better, synthesized speech and intonation perception. Our model includes a conditioning mechanism that effectively handles this issue by allowing emotions to contribute to each phone with varying intensity levels. The human assessment indicates that proposed modifications generate audio with higher MOS and emotional expressiveness.

State of the art (SOTA) neural text to speech (TTS) models can generate natural-sounding synthetic voices. These models are characterized by large memory footprints and substantial number of operations due to the long-standing focus on speech quality with cloud inference in mind. Neural TTS models are generally not designed to perform standalone speech syntheses on resource-constrained and no Internet access edge devices. In this work, an efficient neural TTS called EfficientSpeech that synthesizes speech on an ARM CPU in real-time is proposed. EfficientSpeech uses a shallow non-autoregressive pyramid-structure transformer forming a U-Network. EfficientSpeech has 266k parameters and consumes 90 MFLOPS only or about 1% of the size and amount of computation in modern compact models such as Mixer-TTS. EfficientSpeech achieves an average mel generation real-time factor of 104.3 on an RPi4. Human evaluation shows only a slight degradation in audio quality as compared to FastSpeech2.

Disclaimer: This post may be slightly dryer and technical than my usual posts, but I'll try my best to simplify and make it less painful to read. Conventional wisdom tells us when it comes to Machine/Deep Learning you need large quantity of data. While this is true, in situations where large quantity of relevant data is not readily available, it shouldn't prevent you from pursuing to solve a problem using Deep Learning. It should be evaluated on a per-case basis. Don't just rule out adopting Machine/Deep Learning because "there is not enough data".