Ke Hu

Yifan Peng, Krishna C. Puvvada, Zhehuai Chen et al. · nvidia

Recent studies have augmented large language models (LLMs) with speech capabilities, leading to the development of speech language models (SpeechLMs). Earlier SpeechLMs focused on single-turn speech-based question answering (QA), where user input comprised a speech context and a text question. More recent studies have extended this to multi-turn conversations, though they often require complex, multi-stage supervised fine-tuning (SFT) with diverse data. Another critical challenge with SpeechLMs is catastrophic forgetting, where models optimized for speech tasks suffer significant degradation in text-only performance. To mitigate these issues, we propose a novel single-stage joint speech-text SFT approach on the low-rank adaptation (LoRA) of the LLM backbone. Our joint SFT combines text-only SFT data with three types of speech-related data: speech recognition and translation, speech-based QA, and mixed-modal SFT. Compared to previous SpeechLMs with 7B or 13B parameters, our 3B model demonstrates superior performance across various speech benchmarks while preserving the original capabilities on text-only tasks. Furthermore, our model shows emergent abilities of effectively handling previously unseen prompts and tasks, including multi-turn, mixed-modal inputs.

Cal Peyser, Zhong Meng, Ke Hu et al.

The last year has seen astonishing progress in text-prompted image generation premised on the idea of a cross-modal representation space in which the text and image domains are represented jointly. In ASR, this idea has found application as joint speech-text encoders that can scale to the capacities of very large parameter models by being trained on both unpaired speech and text. While these methods show promise, they have required special treatment of the sequence-length mismatch inherent in speech and text, either by up-sampling heuristics or an explicit alignment model. In this work, we offer evidence that joint speech-text encoders naturally achieve consistent representations across modalities by disregarding sequence length, and argue that consistency losses could forgive length differences and simply assume the best alignment. We show that such a loss improves downstream WER in both a large-parameter monolingual and multilingual system.

Ke Hu, Zhehuai Chen, Chao-Han Huck Yang et al.

Large language models (LLMs) have demonstrated remarkable advancements in language understanding and generation. Building on the success of text-based LLMs, recent research has adapted these models to use speech embeddings for prompting, resulting in Speech-LLM models that exhibit strong performance in automatic speech recognition (ASR) and automatic speech translation (AST). In this work, we propose a novel approach to leverage ASR transcripts as prompts for AST in a Speech-LLM built on an encoder-decoder text LLM. The Speech-LLM model consists of a speech encoder and an encoder-decoder structure Megatron-T5. By first decoding speech to generate ASR transcripts and subsequently using these transcripts along with encoded speech for prompting, we guide the speech translation in a two-step process like chain-of-thought (CoT) prompting. Low-rank adaptation (LoRA) is used for the T5 LLM for model adaptation and shows superior performance to full model fine-tuning. Experimental results show that the proposed CoT prompting significantly improves AST performance, achieving an average increase of 2.4 BLEU points across 6 En->X or X->En AST tasks compared to speech prompting alone. Additionally, compared to a related CoT prediction method that predicts a concatenated sequence of ASR and AST transcripts, our method performs better by an average of 2 BLEU points.

Piotr Żelasko, Zhehuai Chen, Mengru Wang et al.

A rising interest in the modality extension of foundation language models warrants discussion on the most effective, and efficient, multimodal training approach. This work focuses on neural machine translation (NMT) and proposes a joint multimodal training regime of Speech-LLM to include automatic speech translation (AST). We investigate two different foundation model architectures, decoder-only GPT and encoder-decoder T5, extended with Canary-1B's speech encoder. To handle joint multimodal training, we propose a novel training framework called EMMeTT. EMMeTT improves training efficiency with the following: balanced sampling across languages, datasets, and modalities; efficient sequential data iteration; and a novel 2D bucketing scheme for multimodal data, complemented by a batch size optimizer (OOMptimizer). We show that a multimodal training consistently helps with both architectures. Moreover, SALM-T5 trained with EMMeTT retains the original NMT capability while outperforming AST baselines on four-language subsets of FLORES and FLEURS. The resultant Multimodal Translation Model produces strong text and speech translation results at the same time.

Xin Chen, Qingtao Tang, Ke Hu et al.

User historical behaviors are proved useful for Click Through Rate (CTR) prediction in online advertising system. In Meituan, one of the largest e-commerce platform in China, an item is typically displayed with its image and whether a user clicks the item or not is usually influenced by its image, which implies that user's image behaviors are helpful for understanding user's visual preference and improving the accuracy of CTR prediction. Existing user image behavior models typically use a two-stage architecture, which extracts visual embeddings of images through off-the-shelf Convolutional Neural Networks (CNNs) in the first stage, and then jointly trains a CTR model with those visual embeddings and non-visual features. We find that the two-stage architecture is sub-optimal for CTR prediction. Meanwhile, precisely labeled categories in online ad systems contain abundant visual prior information, which can enhance the modeling of user image behaviors. However, off-the-shelf CNNs without category prior may extract category unrelated features, limiting CNN's expression ability. To address the two issues, we propose a hybrid CNN based attention module, unifying user's image behaviors and category prior, for CTR prediction. Our approach achieves significant improvements in both online and offline experiments on a billion scale real serving dataset.

W. Ronny Huang, Cyril Allauzen, Tongzhou Chen et al.

In the era of large models, the autoregressive nature of decoding often results in latency serving as a significant bottleneck. We propose a non-autoregressive LM-fused ASR system that effectively leverages the parallelization capabilities of accelerator hardware. Our approach combines the Universal Speech Model (USM) and the PaLM 2 language model in per-segment scoring mode, achieving an average relative WER improvement across all languages of 10.8% on FLEURS and 3.6% on YouTube captioning. Furthermore, our comprehensive ablation study analyzes key parameters such as LLM size, context length, vocabulary size, fusion methodology. For instance, we explore the impact of LLM size ranging from 128M to 340B parameters on ASR performance. This study provides valuable insights into the factors influencing the effectiveness of practical large-scale LM-fused speech recognition systems.

Bolin Ni, Hongbo Zhao, Chenghao Zhang et al.

Continual learning (CL) aims to empower models to learn new tasks without forgetting previously acquired knowledge. Most prior works concentrate on the techniques of architectures, replay data, regularization, \etc. However, the category name of each class is largely neglected. Existing methods commonly utilize the one-hot labels and randomly initialize the classifier head. We argue that the scarce semantic information conveyed by the one-hot labels hampers the effective knowledge transfer across tasks. In this paper, we revisit the role of the classifier head within the CL paradigm and replace the classifier with semantic knowledge from pretrained language models (PLMs). Specifically, we use PLMs to generate semantic targets for each class, which are frozen and serve as supervision signals during training. Such targets fully consider the semantic correlation between all classes across tasks. Empirical studies show that our approach mitigates forgetting by alleviating representation drifting and facilitating knowledge transfer across tasks. The proposed method is simple to implement and can seamlessly be plugged into existing methods with negligible adjustments. Extensive experiments based on eleven mainstream baselines demonstrate the effectiveness and generalizability of our approach to various protocols. For example, under the class-incremental learning setting on ImageNet-100, our method significantly improves the Top-1 accuracy by 3.2\% to 6.1\% while reducing the forgetting rate by 2.6\% to 13.1\%.

Ke Hu, Ehsan Hosseini-Asl, Chen Chen et al. · nvidia

Spoken dialogue is an intuitive form of human-computer interaction, yet current speech language models often remain constrained to turn-based exchanges, lacking real-time adaptability such as user barge-in. We propose a novel duplex speech to speech (S2S) architecture featuring continuous user inputs and codec agent outputs with channel fusion that directly models simultaneous user and agent streams. Using a pretrained streaming encoder for user input enables the first duplex S2S model without requiring speech pretrain. Separate architectures for agent and user modeling facilitate codec fine-tuning for better agent voices and halve the bitrate (0.6 kbps) compared to previous works. Experimental results show that the proposed model outperforms previous duplex models in reasoning, turn-taking, and barge-in abilities. The model requires significantly less speech data, as speech pretrain is skipped, which markedly simplifies the process of building a duplex S2S model from any LLMs. Finally, it is the first openly available duplex S2S model with training and inference code to foster reproducibility.

Yen-Ting Lin, Chao-Han Huck Yang, Zhehuai Chen et al.

Construction of a general-purpose post-recognition error corrector poses a crucial question: how can we most effectively train a model on a large mixture of domain datasets? The answer would lie in learning dataset-specific features and digesting their knowledge in a single model. Previous methods achieve this by having separate correction language models, resulting in a significant increase in parameters. In this work, we present Mixture-of-Experts as a solution, highlighting that MoEs are much more than a scalability tool. We propose a Multi-Task Correction MoE, where we train the experts to become an ``expert'' of speech-to-text, language-to-text and vision-to-text datasets by learning to route each dataset's tokens to its mapped expert. Experiments on the Open ASR Leaderboard show that we explore a new state-of-the-art performance by achieving an average relative $5.0$% WER reduction and substantial improvements in BLEU scores for speech and translation tasks. On zero-shot evaluation, NeKo outperforms GPT-3.5 and Claude-Opus with $15.5$% to $27.6$% relative WER reduction in the Hyporadise benchmark. NeKo performs competitively on grammar and post-OCR correction as a multi-task model.

Ke Hu, Krishna Puvvada, Elena Rastorgueva et al.

We introduce a data-driven approach for enabling word-level timestamp prediction in the Canary model. Accurate timestamp information is crucial for a variety of downstream tasks such as speech content retrieval and timed subtitles. While traditional hybrid systems and end-to-end (E2E) models may employ external modules for timestamp prediction, our approach eliminates the need for separate alignment mechanisms. By leveraging the NeMo Forced Aligner (NFA) as a teacher model, we generate word-level timestamps and train the Canary model to predict timestamps directly. We introduce a new <|timestamp|> token, enabling the Canary model to predict start and end timestamps for each word. Our method demonstrates precision and recall rates between 80% and 90%, with timestamp prediction errors ranging from 20 to 120 ms across four languages, with minimal WER degradation. Additionally, we extend our system to automatic speech translation (AST) tasks, achieving timestamp prediction errors around 200 milliseconds.

Zhen Wan, Chao-Han Huck Yang, Yahan Yu et al.

We introduce Speech-based Intelligence Quotient (SIQ) as a new form of human cognition-inspired evaluation pipeline for voice understanding large language models, LLM Voice, designed to assess their voice understanding ability. Moving beyond popular voice understanding metrics such as word error rate (WER), SIQ examines LLM Voice across three cognitive levels motivated by Bloom's Taxonomy: (1) Remembering (i.e., WER for verbatim accuracy); (2) Understanding (i.e., similarity of LLM's interpretations); and (3) Application (i.e., QA accuracy for simulating downstream tasks). We demonstrate that SIQ not only quantifies voice understanding abilities but also provides unified comparisons between cascaded methods (e.g., ASR LLM) and end-to-end models, identifies annotation errors in existing benchmarks, and detects hallucinations in LLM Voice. Our framework represents a first-of-its-kind intelligence examination that bridges cognitive principles with voice-oriented benchmarks, while exposing overlooked challenges in multi-modal training.

Ke Hu, WeiDong Qiu, Peng Tang

In the field of federated learning, addressing non-independent and identically distributed (non-i.i.d.) data remains a quintessential challenge for improving global model performance. This work introduces the Feature Norm Regularized Federated Learning (FNR-FL) algorithm, which uniquely incorporates class average feature norms to enhance model accuracy and convergence in non-i.i.d. scenarios. Our comprehensive analysis reveals that FNR-FL not only accelerates convergence but also significantly surpasses other contemporary federated learning algorithms in test accuracy, particularly under feature distribution skew scenarios. The novel modular design of FNR-FL facilitates seamless integration with existing federated learning frameworks, reinforcing its adaptability and potential for widespread application. We substantiate our claims through rigorous empirical evaluations, demonstrating FNR-FL's exceptional performance across various skewed data distributions. Relative to FedAvg, FNR-FL exhibits a substantial 66.24\% improvement in accuracy and a significant 11.40\% reduction in training time, underscoring its enhanced effectiveness and efficiency.

Ke Hu, Bo Li, Tara N. Sainath et al.

End-to-end models with large capacity have significantly improved multilingual automatic speech recognition, but their computation cost poses challenges for on-device applications. We propose a streaming truly multilingual Conformer incorporating mixture-of-expert (MoE) layers that learn to only activate a subset of parameters in training and inference. The MoE layer consists of a softmax gate which chooses the best two experts among many in forward propagation. The proposed MoE layer offers efficient inference by activating a fixed number of parameters as the number of experts increases. We evaluate the proposed model on a set of 12 languages, and achieve an average 11.9% relative improvement in WER over the baseline. Compared to an adapter model using ground truth information, our MoE model achieves similar WER and activates similar number of parameters but without any language information. We further show around 3% relative WER improvement by multilingual shallow fusion.

Jianqiang Huang, Ke Hu, Qingtao Tang et al.

Click-through rate (CTR) prediction plays an important role in online advertising and recommender systems. In practice, the training of CTR models depends on click data which is intrinsically biased towards higher positions since higher position has higher CTR by nature. Existing methods such as actual position training with fixed position inference and inverse propensity weighted training with no position inference alleviate the bias problem to some extend. However, the different treatment of position information between training and inference will inevitably lead to inconsistency and sub-optimal online performance. Meanwhile, the basic assumption of these methods, i.e., the click probability is the product of examination probability and relevance probability, is oversimplified and insufficient to model the rich interaction between position and other information. In this paper, we propose a Deep Position-wise Interaction Network (DPIN) to efficiently combine all candidate items and positions for estimating CTR at each position, achieving consistency between offline and online as well as modeling the deep non-linear interaction among position, user, context and item under the limit of serving performance. Following our new treatment to the position bias in CTR prediction, we propose a new evaluation metrics named PAUC (position-wise AUC) that is suitable for measuring the ranking quality at a given position. Through extensive experiments on a real world dataset, we show empirically that our method is both effective and efficient in solving position bias problem. We have also deployed our method in production and observed statistically significant improvement over a highly optimized baseline in a rigorous A/B test.

David Qiu, Qiujia Li, Yanzhang He et al.

We study the problem of word-level confidence estimation in subword-based end-to-end (E2E) models for automatic speech recognition (ASR). Although prior works have proposed training auxiliary confidence models for ASR systems, they do not extend naturally to systems that operate on word-pieces (WP) as their vocabulary. In particular, ground truth WP correctness labels are needed for training confidence models, but the non-unique tokenization from word to WP causes inaccurate labels to be generated. This paper proposes and studies two confidence models of increasing complexity to solve this problem. The final model uses self-attention to directly learn word-level confidence without needing subword tokenization, and exploits full context features from multiple hypotheses to improve confidence accuracy. Experiments on Voice Search and long-tail test sets show standard metrics (e.g., NCE, AUC, RMSE) improving substantially. The proposed confidence module also enables a model selection approach to combine an on-device E2E model with a hybrid model on the server to address the rare word recognition problem for the E2E model.

Ke Hu, Ruoming Pang, Tara N. Sainath et al.

Interactive speech recognition systems must generate words quickly while also producing accurate results. Two-pass models excel at these requirements by employing a first-pass decoder that quickly emits words, and a second-pass decoder that requires more context but is more accurate. Previous work has established that a deliberation network can be an effective second-pass model. The model attends to two kinds of inputs at once: encoded audio frames and the hypothesis text from the first-pass model. In this work, we explore using transformer layers instead of long-short term memory (LSTM) layers for deliberation rescoring. In transformer layers, we generalize the "encoder-decoder" attention to attend to both encoded audio and first-pass text hypotheses. The output context vectors are then combined by a merger layer. Compared to LSTM-based deliberation, our best transformer deliberation achieves 7% relative word error rate improvements along with a 38% reduction in computation. We also compare against non-deliberation transformer rescoring, and find a 9% relative improvement.

Tara N. Sainath, Yanzhang He, Bo Li et al.

Thus far, end-to-end (E2E) models have not been shown to outperform state-of-the-art conventional models with respect to both quality, i.e., word error rate (WER), and latency, i.e., the time the hypothesis is finalized after the user stops speaking. In this paper, we develop a first-pass Recurrent Neural Network Transducer (RNN-T) model and a second-pass Listen, Attend, Spell (LAS) rescorer that surpasses a conventional model in both quality and latency. On the quality side, we incorporate a large number of utterances across varied domains to increase acoustic diversity and the vocabulary seen by the model. We also train with accented English speech to make the model more robust to different pronunciations. In addition, given the increased amount of training data, we explore a varied learning rate schedule. On the latency front, we explore using the end-of-sentence decision emitted by the RNN-T model to close the microphone, and also introduce various optimizations to improve the speed of LAS rescoring. Overall, we find that RNN-T+LAS offers a better WER and latency tradeoff compared to a conventional model. For example, for the same latency, RNN-T+LAS obtains a 8% relative improvement in WER, while being more than 400-times smaller in model size.

Ke Hu, Tara N. Sainath, Ruoming Pang et al.

End-to-end (E2E) models have made rapid progress in automatic speech recognition (ASR) and perform competitively relative to conventional models. To further improve the quality, a two-pass model has been proposed to rescore streamed hypotheses using the non-streaming Listen, Attend and Spell (LAS) model while maintaining a reasonable latency. The model attends to acoustics to rescore hypotheses, as opposed to a class of neural correction models that use only first-pass text hypotheses. In this work, we propose to attend to both acoustics and first-pass hypotheses using a deliberation network. A bidirectional encoder is used to extract context information from first-pass hypotheses. The proposed deliberation model achieves 12% relative WER reduction compared to LAS rescoring in Google Voice Search (VS) tasks, and 23% reduction on a proper noun test set. Compared to a large conventional model, our best model performs 21% relatively better for VS. In terms of computational complexity, the deliberation decoder has a larger size than the LAS decoder, and hence requires more computations in second-pass decoding.

Ke Hu, Antoine Bruguier, Tara N. Sainath et al.

Contextual automatic speech recognition, i.e., biasing recognition towards a given context (e.g. user's playlists, or contacts), is challenging in end-to-end (E2E) models. Such models maintain a limited number of candidates during beam-search decoding, and have been found to recognize rare named entities poorly. The problem is exacerbated when biasing towards proper nouns in foreign languages, e.g., geographic location names, which are virtually unseen in training and are thus out-of-vocabulary (OOV). While grapheme or wordpiece E2E models might have a difficult time spelling OOV words, phonemes are more acoustically salient and past work has shown that E2E phoneme models can better predict such words. In this work, we propose an E2E model containing both English wordpieces and phonemes in the modeling space, and perform contextual biasing of foreign words at the phoneme level by mapping pronunciations of foreign words into similar English phonemes. In experimental evaluations, we find that the proposed approach performs 16% better than a grapheme-only biasing model, and 8% better than a wordpiece-only biasing model on a foreign place name recognition task, with only slight degradation on regular English tasks.

Antonio Toral, Sheila Castilho, Ke Hu et al.

We reassess a recent study (Hassan et al., 2018) that claimed that machine translation (MT) has reached human parity for the translation of news from Chinese into English, using pairwise ranking and considering three variables that were not taken into account in that previous study: the language in which the source side of the test set was originally written, the translation proficiency of the evaluators, and the provision of inter-sentential context. If we consider only original source text (i.e. not translated from another language, or translationese), then we find evidence showing that human parity has not been achieved. We compare the judgments of professional translators against those of non-experts and discover that those of the experts result in higher inter-annotator agreement and better discrimination between human and machine translations. In addition, we analyse the human translations of the test set and identify important translation issues. Finally, based on these findings, we provide a set of recommendations for future human evaluations of MT.