Ruchao Fan

Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson et al. · microsoft-research

We introduce Phi-4-Mini and Phi-4-Multimodal, compact yet highly capable language and multimodal models. Phi-4-Mini is a 3.8-billion-parameter language model trained on high-quality web and synthetic data, significantly outperforming recent open-source models of similar size and matching the performance of models twice its size on math and coding tasks requiring complex reasoning. This achievement is driven by a carefully curated synthetic data recipe emphasizing high-quality math and coding datasets. Compared to its predecessor, Phi-3.5-Mini, Phi-4-Mini features an expanded vocabulary size of 200K tokens to better support multilingual applications, as well as group query attention for more efficient long-sequence generation. Phi-4-Multimodal is a multimodal model that integrates text, vision, and speech/audio input modalities into a single model. Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. For example, it now ranks first in the OpenASR leaderboard to date, although the LoRA component of the speech/audio modality has just 460 million parameters. Phi-4-Multimodal supports scenarios involving (vision + language), (vision + speech), and (speech/audio) inputs, outperforming larger vision-language and speech-language models on a wide range of tasks. Additionally, we experiment to further train Phi-4-Mini to enhance its reasoning capabilities. Despite its compact 3.8-billion-parameter size, this experimental version achieves reasoning performance on par with or surpassing significantly larger models, including DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B.

Keqi Deng, Shaoshi Ling, Ruchao Fan et al.

Long-context inference in large language models (LLMs) is bottlenecked by the linear growth of the self-attention key-value (KV) cache. Top-k sparse attention alleviates this by loading only a small fraction of the KV cache, but accurately and cheaply estimating cache importance, for both training-free use and sparsity-aware training, remains challenging. This paper proposes UNIQUE, a universal top-k sparse attention framework that addresses both requirements and stays consistently effective across LLM modalities. UNIQUE operates at the granularity of KV pages and estimates per-page importance with a simple yet accurate score combining the mean of the page's keys as a representative vector with their standard deviation as an offset term. To further close the train-inference gap, this paper introduces a soft-mask sparsity-aware training scheme that uses the top-k score boundary as a per-query threshold and a sigmoid soft mask around it, requiring neither auxiliary losses nor architectural changes. Experiments on text and speech LLMs show that UNIQUE preserves task performance on long-context benchmarks such as LongBench Pro and on long-form speech recognition, while delivering up to 11.4x attention-kernel speedup over FlashInfer dense attention and at least 5.3x end-to-end decoding speedup over a vLLM-based dense model.

Ruchao Fan, Wei Chu, Peng Chang et al.

Recently, end-to-end models have been widely used in automatic speech recognition (ASR) systems. Two of the most representative approaches are connectionist temporal classification (CTC) and attention-based encoder-decoder (AED) models. Autoregressive transformers, variants of AED, adopt an autoregressive mechanism for token generation and thus are relatively slow during inference. In this paper, we present a comprehensive study of a CTC Alignment-based Single-Step Non-Autoregressive Transformer (CASS-NAT) for end-to-end ASR. In CASS-NAT, word embeddings in the autoregressive transformer (AT) are substituted with token-level acoustic embeddings (TAE) that are extracted from encoder outputs with the acoustical boundary information offered by the CTC alignment. TAE can be obtained in parallel, resulting in a parallel generation of output tokens. During training, Viterbi-alignment is used for TAE generation, and multiple training strategies are further explored to improve the word error rate (WER) performance. During inference, an error-based alignment sampling method is investigated in depth to reduce the alignment mismatch in the training and testing processes. Experimental results show that the CASS-NAT has a WER that is close to AT on various ASR tasks, while providing a ~24x inference speedup. With and without self-supervised learning, we achieve new state-of-the-art results for non-autoregressive models on several datasets. We also analyze the behavior of the CASS-NAT decoder to explain why it can perform similarly to AT. We find that TAEs have similar functionality to word embeddings for grammatical structures, which might indicate the possibility of learning some semantic information from TAEs without a language model.

Keqi Deng, Ruchao Fan, Bo Ren et al.

Despite extensions to speech inputs, effectively leveraging the rich knowledge and contextual understanding of large language models (LLMs) in automatic speech recognition (ASR) remains non-trivial, as the task primarily involves direct speech-to-text mapping. To address this, this paper proposes chain-of-thought ASR (CoT-ASR), which constructs a reasoning chain that enables LLMs to first analyze the input speech and generate contextual analysis, thereby fully exploiting their generative capabilities. With this contextual reasoning, CoT-ASR then performs more informed speech recognition and completes both reasoning and transcription in a single pass. Moreover, CoT-ASR naturally supports user-guided transcription: while designed to self-generate reasoning, it can also seamlessly incorporate user-provided context to guide transcription, further extending ASR functionality. To reduce the modality gap, this paper introduces a CTC-guided Modality Adapter, which uses CTC non-blank token probabilities to weight LLM embeddings, efficiently aligning speech encoder outputs with the LLM's textual latent space. Experiments show that, compared to standard LLM-based ASR, CoT-ASR achieves a relative reduction of 8.7% in word error rate (WER) and 16.9% in entity error rate (EER).

Ruchao Fan, Natarajan Balaji Shanka, Abeer Alwan

Non-autoregressive automatic speech recognition (NASR) models have gained attention due to their parallelism and fast inference. The encoder-based NASR, e.g. connectionist temporal classification (CTC), can be initialized from the speech foundation models (SFM) but does not account for any dependencies among intermediate tokens. The encoder-decoder-based NASR, like CTC alignment-based single-step non-autoregressive transformer (CASS-NAT), can mitigate the dependency problem but is not able to efficiently integrate SFM. Inspired by the success of recent work of speech-text joint pre-training with a shared transformer encoder, we propose a new encoder-based NASR, UniEnc-CASSNAT, to combine the advantages of CTC and CASS-NAT. UniEnc-CASSNAT consists of only an encoder as the major module, which can be the SFM. The encoder plays the role of both the CASS-NAT encoder and decoder by two forward passes. The first pass of the encoder accepts the speech signal as input, while the concatenation of the speech signal and the token-level acoustic embedding is used as the input for the second pass. Examined on the Librispeech 100h, MyST, and Aishell1 datasets, the proposed UniEnc-CASSNAT achieves state-of-the-art NASR results and is better or comparable to CASS-NAT with only an encoder and hence, fewer model parameters. Our codes are publicly available.

Wei Chu, Yuanzhe Dong, Ke Tan et al.

OleSpeech-IV dataset is a large-scale multispeaker and multilingual conversational speech dataset with diverse topics. The audio content comes from publicly-available English podcasts, talk shows, teleconferences, and other conversations. Speaker names, turns, and transcripts are human-sourced and refined by a proprietary pipeline, while additional information such as timestamps and confidence scores is derived from the pipeline. The IV denotes its position as Tier IV in the Olewave dataset series. In addition, we have open-sourced a subset, OleSpeech-IV-2025-EN-AR-100, for non-commercial research use.

Haibin Wu, Yuxuan Hu, Ruchao Fan et al.

Speech language models (Speech LMs) enable end-to-end speech-text modelling within a single model, offering a promising direction for spoken dialogue systems. The choice of speech-text jointly decoding paradigm plays a critical role in performance, efficiency, and alignment quality. In this work, we systematically compare representative joint speech-text decoding strategies-including the interleaved, and parallel generation paradigms-under a controlled experimental setup using the same base language model, speech tokenizer and training data. Our results show that the interleaved approach achieves the best alignment. However it suffers from slow inference due to long token sequence length. To address this, we propose a novel early-stop interleaved (ESI) pattern that not only significantly accelerates decoding but also yields slightly better performance. Additionally, we curate high-quality question answering (QA) datasets to further improve speech QA performance.

Ruchao Fan, Natarajan Balaji Shankar, Abeer Alwan

Speech foundation models (SFMs) have achieved state-of-the-art results for various speech tasks in supervised (e.g. Whisper) or self-supervised systems (e.g. WavLM). However, the performance of SFMs for child ASR has not been systematically studied. In addition, there is no benchmark for child ASR with standard evaluations, making the comparisons of novel ideas difficult. In this paper, we initiate and present a comprehensive benchmark on several child speech databases based on various SFMs (Whisper, Wav2vec2.0, HuBERT, and WavLM). Moreover, we investigate finetuning strategies by comparing various data augmentation and parameter-efficient finetuning (PEFT) methods. We observe that the behaviors of these methods are different when the model size increases. For example, PEFT matches the performance of full finetuning for large models but worse for small models. To stabilize finetuning using augmented data, we propose a perturbation invariant finetuning (PIF) loss as a regularization.

Yunzheng Zhu, Ruchao Fan, Abeer Alwan

Children's automatic speech recognition (ASR) is always difficult due to, in part, the data scarcity problem, especially for kindergarten-aged kids. When data are scarce, the model might overfit to the training data, and hence good starting points for training are essential. Recently, meta-learning was proposed to learn model initialization (MI) for ASR tasks of different languages. This method leads to good performance when the model is adapted to an unseen language. However, MI is vulnerable to overfitting on training tasks (learner overfitting). It is also unknown whether MI generalizes to other low-resource tasks. In this paper, we validate the effectiveness of MI in children's ASR and attempt to alleviate the problem of learner overfitting. To achieve model-agnostic meta-learning (MAML), we regard children's speech at each age as a different task. In terms of learner overfitting, we propose a task-level augmentation method by simulating new ages using frequency warping techniques. Detailed experiments are conducted to show the impact of task augmentation on each age for kindergarten-aged speech. As a result, our approach achieves a relative word error rate (WER) improvement of 51% over the baseline system with no augmentation or initialization.

Alexander Johnson, Ruchao Fan, Robin Morris et al.

This paper proposes a novel linear prediction coding-based data aug-mentation method for children's low and zero resource dialect ASR. The data augmentation procedure consists of perturbing the formant peaks of the LPC spectrum during LPC analysis and reconstruction. The method is evaluated on two novel children's speech datasets with one containing California English from the Southern CaliforniaArea and the other containing a mix of Southern American English and African American English from the Atlanta, Georgia area. We test the proposed method in training both an HMM-DNN system and an end-to-end system to show model-robustness and demonstrate that the algorithm improves ASR performance, especially for zero resource dialect children's task, as compared to common data augmentation methods such as VTLP, Speed Perturbation, and SpecAugment.

Jinhan Wang, Yunzheng Zhu, Ruchao Fan et al.

This paper describes the SPAPL system for the INTERSPEECH 2021 Challenge: Shared Task on Automatic Speech Recognition for Non-Native Children's Speech in German. ~ 5 hours of transcribed data and ~ 60 hours of untranscribed data are provided to develop a German ASR system for children. For the training of the transcribed data, we propose a non-speech state discriminative loss (NSDL) to mitigate the influence of long-duration non-speech segments within speech utterances. In order to explore the use of the untranscribed data, various approaches are implemented and combined together to incrementally improve the system performance. First, bidirectional autoregressive predictive coding (Bi-APC) is used to learn initial parameters for acoustic modelling using the provided untranscribed data. Second, incremental semi-supervised learning is further used to iteratively generate pseudo-transcribed data. Third, different data augmentation schemes are used at different training stages to increase the variability and size of the training data. Finally, a recurrent neural network language model (RNNLM) is used for rescoring. Our system achieves a word error rate (WER) of 39.68% on the evaluation data, an approximately 12% relative improvement over the official baseline (45.21%).

Ruchao Fan, Wei Chu, Peng Chang et al.

Non-autoregressive mechanisms can significantly decrease inference time for speech transformers, especially when the single step variant is applied. Previous work on CTC alignment-based single step non-autoregressive transformer (CASS-NAT) has shown a large real time factor (RTF) improvement over autoregressive transformers (AT). In this work, we propose several methods to improve the accuracy of the end-to-end CASS-NAT, followed by performance analyses. First, convolution augmented self-attention blocks are applied to both the encoder and decoder modules. Second, we propose to expand the trigger mask (acoustic boundary) for each token to increase the robustness of CTC alignments. In addition, iterated loss functions are used to enhance the gradient update of low-layer parameters. Without using an external language model, the WERs of the improved CASS-NAT, when using the three methods, are 3.1%/7.2% on Librispeech test clean/other sets and the CER is 5.4% on the Aishell1 test set, achieving a 7%~21% relative WER/CER improvement. For the analyses, we plot attention weight distributions in the decoders to visualize the relationships between token-level acoustic embeddings. When the acoustic embeddings are visualized, we find that they have a similar behavior to word embeddings, which explains why the improved CASS-NAT performs similarly to AT.

Gary Yeung, Ruchao Fan, Abeer Alwan

Automatic speech recognition (ASR) systems for young children are needed due to the importance of age-appropriate educational technology. Because of the lack of publicly available young child speech data, feature extraction strategies such as feature normalization and data augmentation must be considered to successfully train child ASR systems. This study proposes a novel technique for child ASR using both feature normalization and data augmentation methods based on the relationship between formants and fundamental frequency ($f_o$). Both the $f_o$ feature normalization and data augmentation techniques are implemented as a frequency shift in the Mel domain. These techniques are evaluated on a child read speech ASR task. Child ASR systems are trained by adapting a BLSTM-based acoustic model trained on adult speech. Using both $f_o$ normalization and data augmentation results in a relative word error rate (WER) improvement of 19.3% over the baseline when tested on the OGI Kids' Speech Corpus, and the resulting child ASR system achieves the best WER currently reported on this corpus.

Ruchao Fan, Amber Afshan, Abeer Alwan

We present a bidirectional unsupervised model pre-training (UPT) method and apply it to children's automatic speech recognition (ASR). An obstacle to improving child ASR is the scarcity of child speech databases. A common approach to alleviate this problem is model pre-training using data from adult speech. Pre-training can be done using supervised (SPT) or unsupervised methods, depending on the availability of annotations. Typically, SPT performs better. In this paper, we focus on UPT to address the situations when pre-training data are unlabeled. Autoregressive predictive coding (APC), a UPT method, predicts frames from only one direction, limiting its use to uni-directional pre-training. Conventional bidirectional UPT methods, however, predict only a small portion of frames. To extend the benefits of APC to bi-directional pre-training, Bi-APC is proposed. We then use adaptation techniques to transfer knowledge learned from adult speech (using the Librispeech corpus) to child speech (OGI Kids corpus). LSTM-based hybrid systems are investigated. For the uni-LSTM structure, APC obtains similar WER improvements to SPT over the baseline. When applied to BLSTM, however, APC is not as competitive as SPT, but our proposed Bi-APC has comparable improvements to SPT.

Ruchao Fan, Wei Chu, Peng Chang et al.

We propose a CTC alignment-based single step non-autoregressive transformer (CASS-NAT) for speech recognition. Specifically, the CTC alignment contains the information of (a) the number of tokens for decoder input, and (b) the time span of acoustics for each token. The information are used to extract acoustic representation for each token in parallel, referred to as token-level acoustic embedding which substitutes the word embedding in autoregressive transformer (AT) to achieve parallel generation in decoder. During inference, an error-based alignment sampling method is proposed to be applied to the CTC output space, reducing the WER and retaining the parallelism as well. Experimental results show that the proposed method achieves WERs of 3.8%/9.1% on Librispeech test clean/other dataset without an external LM, and a CER of 5.8% on Aishell1 Mandarin corpus, respectively1. Compared to the AT baseline, the CASS-NAT has a performance reduction on WER, but is 51.2x faster in terms of RTF. When decoding with an oracle CTC alignment, the lower bound of WER without LM reaches 2.3% on the test-clean set, indicating the potential of the proposed method.

Vijay Ravi, Ruchao Fan, Amber Afshan et al.

In this paper, we propose a novel way of addressing text-dependent automatic speaker verification (TD-ASV) by using a shared-encoder with task-specific decoders. An autoregressive predictive coding (APC) encoder is pre-trained in an unsupervised manner using both out-of-domain (LibriSpeech, VoxCeleb) and in-domain (DeepMine) unlabeled datasets to learn generic, high-level feature representation that encapsulates speaker and phonetic content. Two task-specific decoders were trained using labeled datasets to classify speakers (SID) and phrases (PID). Speaker embeddings extracted from the SID decoder were scored using a PLDA. SID and PID systems were fused at the score level. There is a 51.9% relative improvement in minDCF for our system compared to the fully supervised x-vector baseline on the cross-lingual DeepMine dataset. However, the i-vector/HMM method outperformed the proposed APC encoder-decoder system. A fusion of the x-vector/PLDA baseline and the SID/PLDA scores prior to PID fusion further improved performance by 15% indicating complementarity of the proposed approach to the x-vector system. We show that the proposed approach can leverage from large, unlabeled, data-rich domains, and learn speech patterns independent of downstream tasks. Such a system can provide competitive performance in domain-mismatched scenarios where test data is from data-scarce domains.

Pan Zhou, Ruchao Fan, Wei Chen et al.

Transformer has shown promising results in many sequence to sequence transformation tasks recently. It utilizes a number of feed-forward self-attention layers to replace the recurrent neural networks (RNN) in attention-based encoder decoder (AED) architecture. Self-attention layer learns temporal dependence by incorporating sinusoidal positional embedding of tokens in a sequence for parallel computing. Quicker iteration speed in training than sequential operation of RNN can be obtained. Deeper layers of the transformer also make it perform better than RNN-based AED. However, this parallelization ability is lost when applying scheduled sampling training. Self-attention with sinusoidal positional embedding may cause performance degradations for longer sequences that have similar acoustic or semantic information at different positions as well. To address these problems, we propose to use parallel scheduled sampling (PSS) and relative positional embedding (RPE) to help the transformer generalize to unseen data. Our proposed methods achieve a 7% relative improvement for short utterances and a 70% relative gain for long utterances on a 10,000-hour Mandarin ASR task.

Ruchao Fan, Pan Zhou, Wei Chen et al.

Attention-based end-to-end models such as Listen, Attend and Spell (LAS), simplify the whole pipeline of traditional automatic speech recognition (ASR) systems and become popular in the field of speech recognition. In previous work, researchers have shown that such architectures can acquire comparable results to state-of-the-art ASR systems, especially when using a bidirectional encoder and global soft attention (GSA) mechanism. However, bidirectional encoder and GSA are two obstacles for real-time speech recognition. In this work, we aim to stream LAS baseline by removing the above two obstacles. On the encoder side, we use a latency-controlled (LC) bidirectional structure to reduce the delay of forward computation. Meanwhile, an adaptive monotonic chunk-wise attention (AMoChA) mechanism is proposed to replace GSA for the calculation of attention weight distribution. Furthermore, we propose two methods to alleviate the huge performance degradation when combining LC and AMoChA. Finally, we successfully acquire an online LAS model, LC-AMoChA, which has only 3.5% relative performance reduction to LAS baseline on our internal Mandarin corpus.