Ron Hoory

Avihu Dekel, Slava Shechtman, Raul Fernandez et al.

Large Language Models (LLMs) demonstrate impressive capabilities, yet interaction with these models is mostly facilitated through text. Using Text-To-Speech to synthesize LLM outputs typically results in notable latency, which is impractical for fluent voice conversations. We propose LLM2Speech, an architecture to synthesize speech while text is being generated by an LLM which yields significant latency reduction. LLM2Speech mimics the predictions of a non-streaming teacher model while limiting the exposure to future context in order to enable streaming. It exploits the hidden embeddings of the LLM, a by-product of the text generation that contains informative semantic context. Experimental results show that LLM2Speech maintains the teacher's quality while reducing the latency to enable natural conversations.

Arnon Turetzky, Avihu Dekel, Hagai Aronowitz et al.

Spoken meaning often depends not only on what is said, but also on which word is emphasized. The same sentence can convey correction, contrast, or clarification depending on where emphasis falls. Although modern text-to-speech (TTS) systems generate expressive speech, it remains unclear whether they infer contextually appropriate stress from discourse alone. To address this gap, we present Context-Aware Stress TTS (CAST), a benchmark for evaluating context-conditioned word-level stress in TTS. Items are defined as contrastive context pairs: identical sentences paired with distinct contexts requiring different stressed words. We evaluate state-of-the-art systems and find a consistent gap: text-only language models reliably recover the intended stress from context, yet TTS systems frequently fail to realize it in speech. We release the benchmark, evaluation framework, construction pipeline and a synthetic corpus to support future work on context-aware speech synthesis.

Hagai Aronowitz, Itai Gat, Edmilson Morais et al.

Recent innovations in self-supervised representation learning have led to remarkable advances in natural language processing. That said, in the speech processing domain, self-supervised representation learning-based systems are not yet considered state-of-the-art. We propose leveraging recent advances in self-supervised-based speech processing to create a common speech analysis engine. Such an engine should be able to handle multiple speech processing tasks, using a single architecture, to obtain state-of-the-art accuracy. The engine must also enable support for new tasks with small training datasets. Beyond that, a common engine should be capable of supporting distributed training with client in-house private data. We present the architecture for a common speech analysis engine based on the HuBERT self-supervised speech representation. Based on experiments, we report our results for language identification and emotion recognition on the standard evaluations NIST-LRE 07 and IEMOCAP. Our results surpass the state-of-the-art performance reported so far on these tasks. We also analyzed our engine on the emotion recognition task using reduced amounts of training data and show how to achieve improved results.

Claudio Pinhanez, Raul Fernandez, Marcelo Grave et al. · ibm-research

Representations of AI agents in user interfaces and robotics are predominantly White, not only in terms of facial and skin features, but also in the synthetic voices they use. In this paper we explore some unexpected challenges in the representation of race we found in the process of developing an U.S. English Text-to-Speech (TTS) system aimed to sound like an educated, professional, regional accent-free African American woman. The paper starts by presenting the results of focus groups with African American IT professionals where guidelines and challenges for the creation of a representative and appropriate TTS system were discussed and gathered, followed by a discussion about some of the technical difficulties faced by the TTS system developers. We then describe two studies with U.S. English speakers where the participants were not able to attribute the correct race to the African American TTS voice while overwhelmingly correctly recognizing the race of a White TTS system of similar quality. A focus group with African American IT workers not only confirmed the representativeness of the African American voice we built, but also suggested that the surprising recognition results may have been caused by the inability or the latent prejudice from non-African Americans to associate educated, non-vernacular, professionally-sounding voices to African American people.

Nimrod Shabtay, Zvi Kons, Avihu Dekel et al.

Question answering (QA) systems are designed to answer natural language questions. Visual QA (VQA) and Spoken QA (SQA) systems extend the textual QA system to accept visual and spoken input respectively. This work aims to create a system that enables user interaction through both speech and images. That is achieved through the fusion of text, speech, and image modalities to tackle the task of spoken VQA (SVQA). The resulting multi-modal model has textual, visual, and spoken inputs and can answer spoken questions on images. Training and evaluating SVQA models requires a dataset for all three modalities, but no such dataset currently exists. We address this problem by synthesizing VQA datasets using two zero-shot TTS models. Our initial findings indicate that a model trained only with synthesized speech nearly reaches the performance of the upper-bounding model trained on textual QAs. In addition, we show that the choice of the TTS model has a minor impact on accuracy.

Avishai Elmakies, Hagai Aronowitz, Nimrod Shabtay et al.

In this paper, we introduce a Group Relative Policy Optimization (GRPO)-based method for training Speech-Aware Large Language Models (SALLMs) on open-format speech understanding tasks, such as Spoken Question Answering and Automatic Speech Translation. SALLMs have proven highly effective for speech understanding tasks. GRPO has recently gained traction for its efficiency in training LLMs, and prior work has explored its application to SALLMs, primarily in multiple-choice tasks. Building on this, we focus on open-format tasks that better reflect the generative abilities of the models. Our approach leverages GRPO with BLEU as the reward signal to optimize SALLMs, and we demonstrate empirically that it surpasses standard SFT across several key metrics. Finally, we explore the potential of incorporating off-policy samples within GRPO for these tasks, highlighting avenues for further improvement and further research.

Zvi Kons, Aharon Satt, Hong-Kwang Kuo et al.

Intent classifiers are vital to the successful operation of virtual agent systems. This is especially so in voice activated systems where the data can be noisy with many ambiguous directions for user intents. Before operation begins, these classifiers are generally lacking in real-world training data. Active learning is a common approach used to help label large amounts of collected user input. However, this approach requires many hours of manual labeling work. We present the Nearest Neighbors Scores Improvement (NNSI) algorithm for automatic data selection and labeling. The NNSI reduces the need for manual labeling by automatically selecting highly-ambiguous samples and labeling them with high accuracy. This is done by integrating the classifier's output from a semantically similar group of text samples. The labeled samples can then be added to the training set to improve the accuracy of the classifier. We demonstrated the use of NNSI on two large-scale, real-life voice conversation systems. Evaluation of our results showed that our method was able to select and label useful samples with high accuracy. Adding these new samples to the training data significantly improved the classifiers and reduced error rates by up to 10%.

Edmilson Morais, Ron Hoory, Weizhong Zhu et al.

Self-supervised pre-trained features have consistently delivered state-of-art results in the field of natural language processing (NLP); however, their merits in the field of speech emotion recognition (SER) still need further investigation. In this paper we introduce a modular End-to- End (E2E) SER system based on an Upstream + Downstream architecture paradigm, which allows easy use/integration of a large variety of self-supervised features. Several SER experiments for predicting categorical emotion classes from the IEMOCAP dataset are performed. These experiments investigate interactions among fine-tuning of self-supervised feature models, aggregation of frame-level features into utterance-level features and back-end classification networks. The proposed monomodal speechonly based system not only achieves SOTA results, but also brings light to the possibility of powerful and well finetuned self-supervised acoustic features that reach results similar to the results achieved by SOTA multimodal systems using both Speech and Text modalities.

Itai Gat, Hagai Aronowitz, Weizhong Zhu et al.

Large speech emotion recognition datasets are hard to obtain, and small datasets may contain biases. Deep-net-based classifiers, in turn, are prone to exploit those biases and find shortcuts such as speaker characteristics. These shortcuts usually harm a model's ability to generalize. To address this challenge, we propose a gradient-based adversary learning framework that learns a speech emotion recognition task while normalizing speaker characteristics from the feature representation. We demonstrate the efficacy of our method on both speaker-independent and speaker-dependent settings and obtain new state-of-the-art results on the challenging IEMOCAP dataset.

Samuel Thomas, Hong-Kwang J. Kuo, George Saon et al.

We present a comprehensive study on building and adapting RNN transducer (RNN-T) models for spoken language understanding(SLU). These end-to-end (E2E) models are constructed in three practical settings: a case where verbatim transcripts are available, a constrained case where the only available annotations are SLU labels and their values, and a more restrictive case where transcripts are available but not corresponding audio. We show how RNN-T SLU models can be developed starting from pre-trained automatic speech recognition (ASR) systems, followed by an SLU adaptation step. In settings where real audio data is not available, artificially synthesized speech is used to successfully adapt various SLU models. When evaluated on two SLU data sets, the ATIS corpus and a customer call center data set, the proposed models closely track the performance of other E2E models and achieve state-of-the-art results.

Yinghui Huang, Hong-Kwang Kuo, Samuel Thomas et al.

Training an end-to-end (E2E) neural network speech-to-intent (S2I) system that directly extracts intents from speech requires large amounts of intent-labeled speech data, which is time consuming and expensive to collect. Initializing the S2I model with an ASR model trained on copious speech data can alleviate data sparsity. In this paper, we attempt to leverage NLU text resources. We implemented a CTC-based S2I system that matches the performance of a state-of-the-art, traditional cascaded SLU system. We performed controlled experiments with varying amounts of speech and text training data. When only a tenth of the original data is available, intent classification accuracy degrades by 7.6% absolute. Assuming we have additional text-to-intent data (without speech) available, we investigated two techniques to improve the S2I system: (1) transfer learning, in which acoustic embeddings for intent classification are tied to fine-tuned BERT text embeddings; and (2) data augmentation, in which the text-to-intent data is converted into speech-to-intent data using a multi-speaker text-to-speech system. The proposed approaches recover 80% of performance lost due to using limited intent-labeled speech.

Hong-Kwang J. Kuo, Zoltán Tüske, Samuel Thomas et al.

An essential component of spoken language understanding (SLU) is slot filling: representing the meaning of a spoken utterance using semantic entity labels. In this paper, we develop end-to-end (E2E) spoken language understanding systems that directly convert speech input to semantic entities and investigate if these E2E SLU models can be trained solely on semantic entity annotations without word-for-word transcripts. Training such models is very useful as they can drastically reduce the cost of data collection. We created two types of such speech-to-entities models, a CTC model and an attention-based encoder-decoder model, by adapting models trained originally for speech recognition. Given that our experiments involve speech input, these systems need to recognize both the entity label and words representing the entity value correctly. For our speech-to-entities experiments on the ATIS corpus, both the CTC and attention models showed impressive ability to skip non-entity words: there was little degradation when trained on just entities versus full transcripts. We also explored the scenario where the entities are in an order not necessarily related to spoken order in the utterance. With its ability to do re-ordering, the attention model did remarkably well, achieving only about 2% degradation in speech-to-bag-of-entities F1 score.

Shai Rozenberg, Hagai Aronowitz, Ron Hoory

With the rise of voice-activated applications, the need for speaker recognition is rapidly increasing. The x-vector, an embedding approach based on a deep neural network (DNN), is considered the state-of-the-art when proper end-to-end training is not feasible. However, the accuracy significantly decreases when recording conditions (noise, sample rate, etc.) are mismatched, either between the x-vector training data and the target data or between enrollment and test data. We introduce the Siamese x-vector Reconstruction (SVR) for domain adaptation. We reconstruct the embedding of a higher quality signal from a lower quality counterpart using a lean auxiliary Siamese DNN. We evaluate our method on several mismatch scenarios and demonstrate significant improvement over the baseline.

Zvi Kons, Slava Shechtman, Alex Sorin et al.

We present a lightweight adaptable neural TTS system with high quality output. The system is composed of three separate neural network blocks: prosody prediction, acoustic feature prediction and Linear Prediction Coding Net as a neural vocoder. This system can synthesize speech with close to natural quality while running 3 times faster than real-time on a standard CPU. The modular setup of the system allows for simple adaptation to new voices with a small amount of data. We first demonstrate the ability of the system to produce high quality speech when trained on large, high quality datasets. Following that, we demonstrate its adaptability by mimicking unseen voices using 5 to 20 minutes long datasets with lower recording quality. Large scale Mean Opinion Score quality and similarity tests are presented, showing that the system can adapt to unseen voices with quality gap of 0.12 and similarity gap of 3% compared to natural speech for male voices and quality gap of 0.35 and similarity of gap of 9 % for female voices.