Xingchen Song

Yaoxun Xu, Baiji Liu, Qiaochu Huang and et al.

Due to the mismatch between the source and target domains, how to better utilize the biased word information to improve the performance of the automatic speech recognition model in the target domain becomes a hot research topic. Previous approaches either decode with a fixed external language model or introduce a sizeable biasing module, which leads to poor adaptability and slow inference. In this work, we propose CB-Conformer to improve biased word recognition by introducing the Contextual Biasing Module and the Self-Adaptive Language Model to vanilla Conformer. The Contextual Biasing Module combines audio fragments and contextual information, with only 0.2% model parameters of the original Conformer. The Self-Adaptive Language Model modifies the internal weights of biased words based on their recall and precision, resulting in a greater focus on biased words and more successful integration with the automatic speech recognition model than the standard fixed language model. In addition, we construct and release an open-source Mandarin biased-word dataset based on WenetSpeech. Experiments indicate that our proposed method brings a 15.34% character error rate reduction, a 14.13% biased word recall increase, and a 6.80% biased word F1-score increase compared with the base Conformer.

Dinghao Zhou, Xingchen Song, Di Wu et al.

Continuous audio autoencoders reconstruct waveforms well but often produce latents with weak structure for understanding, while self-supervised audio encoders capture semantics but are not directly decodable. This mismatch complicates a single audio tokenizer that must support both understanding and generation. We adapt continuous autoencoder latents to this setting with two components: a noise-regularized autoencoder bottleneck and a latent-side representation encoder. The bottleneck uses channel normalization and stochastic perturbation instead of KL-based variational training, yielding scale-controlled continuous latents for reconstruction and autoregressive generation. The representation encoder is trained on frozen autoencoder latents with RQ-MTP and frozen-LLM supervision. The resulting tokenizer provides high-dimensional representations for understanding while preserving normalized continuous latents as generation targets

Binbin Zhang, Di Wu, Zhendong Peng et al.

Recently, we made available WeNet, a production-oriented end-to-end speech recognition toolkit, which introduces a unified two-pass (U2) framework and a built-in runtime to address the streaming and non-streaming decoding modes in a single model. To further improve ASR performance and facilitate various production requirements, in this paper, we present WeNet 2.0 with four important updates. (1) We propose U2++, a unified two-pass framework with bidirectional attention decoders, which includes the future contextual information by a right-to-left attention decoder to improve the representative ability of the shared encoder and the performance during the rescoring stage. (2) We introduce an n-gram based language model and a WFST-based decoder into WeNet 2.0, promoting the use of rich text data in production scenarios. (3) We design a unified contextual biasing framework, which leverages user-specific context (e.g., contact lists) to provide rapid adaptation ability for production and improves ASR accuracy in both with-LM and without-LM scenarios. (4) We design a unified IO to support large-scale data for effective model training. In summary, the brand-new WeNet 2.0 achieves up to 10\% relative recognition performance improvement over the original WeNet on various corpora and makes available several important production-oriented features.

Xiaomi LLM-Core Team, Dong Zhang, Gang Wang et al.

Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demonstrates powerful speech continuation capabilities, capable of generating highly realistic talk shows, recitations, livestreaming and debates. At the post-training stage, we curate a diverse instruction-tuning corpus and introduce thinking mechanisms into both audio understanding and generation. MiMo-Audio-7B-Instruct achieves open-source SOTA on audio understanding benchmarks (MMSU, MMAU, MMAR, MMAU-Pro), spoken dialogue benchmarks (Big Bench Audio, MultiChallenge Audio) and instruct-TTS evaluations, approaching or surpassing closed-source models. Model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-Audio.

Yaoxun Xu, Xingchen Song, Zhiyong Wu et al.

In automatic speech recognition, subsampling is essential for tackling diverse scenarios. However, the inadequacy of a single subsampling rate to address various real-world situations often necessitates training and deploying multiple models, consequently increasing associated costs. To address this issue, we propose HydraFormer, comprising HydraSub, a Conformer-based encoder, and a BiTransformer-based decoder. HydraSub encompasses multiple branches, each representing a distinct subsampling rate, allowing for the flexible selection of any branch during inference based on the specific use case. HydraFormer can efficiently manage different subsampling rates, significantly reducing training and deployment expenses. Experiments on AISHELL-1 and LibriSpeech datasets reveal that HydraFormer effectively adapts to various subsampling rates and languages while maintaining high recognition performance. Additionally, HydraFormer showcases exceptional stability, sustaining consistent performance under various initialization conditions, and exhibits robust transferability by learning from pretrained single subsampling rate automatic speech recognition models\footnote{Model code and scripts: https://github.com/HydraFormer/hydraformer}.

Xingchen Song, Di Wu, Zhiyong Wu et al.

In this paper, we present TrimTail, a simple but effective emission regularization method to improve the latency of streaming ASR models. The core idea of TrimTail is to apply length penalty (i.e., by trimming trailing frames, see Fig. 1-(b)) directly on the spectrogram of input utterances, which does not require any alignment. We demonstrate that TrimTail is computationally cheap and can be applied online and optimized with any training loss or any model architecture on any dataset without any extra effort by applying it on various end-to-end streaming ASR networks either trained with CTC loss [1] or Transducer loss [2]. We achieve 100 $\sim$ 200ms latency reduction with equal or even better accuracy on both Aishell-1 and Librispeech. Moreover, by using TrimTail, we can achieve a 400ms algorithmic improvement of User Sensitive Delay (USD) with an accuracy loss of less than 0.2.

Xi Wang, Jie Wang, Xingchen Song et al.

While generative text-to-speech (TTS) models approach human-level quality, monolithic metrics fail to diagnose fine-grained acoustic artifacts or explain perceptual collapse. To address this, we propose TTS-PRISM, a multi-dimensional diagnostic framework for Mandarin. First, we establish a 12-dimensional schema spanning stability to advanced expressiveness. Second, we design a targeted synthesis pipeline with adversarial perturbations and expert anchors to build a high-quality diagnostic dataset. Third, schema-driven instruction tuning embeds explicit scoring criteria and reasoning into an efficient end-to-end model. Experiments on a 1,600-sample Gold Test Set show TTS-PRISM outperforms generalist models in human alignment. Profiling six TTS paradigms establishes intuitive diagnostic flags that reveal fine-grained capability differences. TTS-PRISM is open-source, with code and checkpoints at https://github.com/xiaomi-research/tts-prism.

Xingchen Song, Di Wu, Binbin Zhang et al.

The recently proposed Conformer architecture which combines convolution with attention to capture both local and global dependencies has become the \textit{de facto} backbone model for Automatic Speech Recognition~(ASR). Inherited from the Natural Language Processing (NLP) tasks, the architecture takes Layer Normalization~(LN) as a default normalization technique. However, through a series of systematic studies, we find that LN might take 10\% of the inference time despite that it only contributes to 0.1\% of the FLOPs. This motivates us to replace LN with other normalization techniques, e.g., Batch Normalization~(BN), to speed up inference with the help of operator fusion methods and the avoidance of calculating the mean and variance statistics during inference. After examining several plain attempts which directly remove all LN layers or replace them with BN in the same place, we find that the divergence issue is mainly caused by the unstable layer output. We therefore propose to append a BN layer to each linear or convolution layer where stabilized training results are observed. We also propose to simplify the activations in Conformer, such as Swish and GLU, by replacing them with ReLU. All these exchanged modules can be fused into the weights of the adjacent linear/convolution layers and hence have zero inference cost. Therefore, we name it FusionFormer. Our experiments indicate that FusionFormer is as effective as the LN-based Conformer and is about 10\% faster.

LLM-Core Xiaomi, Bingquan Xia, Bowen Shen et al. · pku

We present MiMo-7B, a large language model born for reasoning tasks, with optimization across both pre-training and post-training stages. During pre-training, we enhance the data preprocessing pipeline and employ a three-stage data mixing strategy to strengthen the base model's reasoning potential. MiMo-7B-Base is pre-trained on 25 trillion tokens, with additional Multi-Token Prediction objective for enhanced performance and accelerated inference speed. During post-training, we curate a dataset of 130K verifiable mathematics and programming problems for reinforcement learning, integrating a test-difficulty-driven code-reward scheme to alleviate sparse-reward issues and employing strategic data resampling to stabilize training. Extensive evaluations show that MiMo-7B-Base possesses exceptional reasoning potential, outperforming even much larger 32B models. The final RL-tuned model, MiMo-7B-RL, achieves superior performance on mathematics, code and general reasoning tasks, surpassing the performance of OpenAI o1-mini. The model checkpoints are available at https://github.com/xiaomimimo/MiMo.

Xingchen Song, Mengtao Xing, Changwei Ma et al.

It is well known that LLM-based systems are data-hungry. Recent LLM-based TTS works typically employ complex data processing pipelines to obtain high-quality training data. These sophisticated pipelines require excellent models at each stage (e.g., speech denoising, speech enhancement, speaker diarization, and punctuation models), which themselves demand high-quality training data and are rarely open-sourced. Even with state-of-the-art models, issues persist, such as incomplete background noise removal and misalignment between punctuation and actual speech pauses. Moreover, the stringent filtering strategies often retain only 10-30\% of the original data, significantly impeding data scaling efforts. In this work, we leverage a noise-robust audio tokenizer (S3Tokenizer) to design a simplified yet effective TTS data processing pipeline that maintains data quality while substantially reducing data acquisition costs, achieving a data retention rate of over 50\%. Beyond data scaling challenges, LLM-based TTS systems also incur higher deployment costs compared to conventional approaches. Current systems typically use LLMs solely for text-to-token generation, while requiring separate models (e.g., flow matching models) for token-to-waveform generation, which cannot be directly executed by LLM inference engines, further complicating deployment. To address these challenges, we eliminate redundant modules in both LLM and flow components, replacing the flow model backbone with an LLM architecture. Building upon this simplified flow backbone, we propose a unified architecture for both streaming and non-streaming inference, significantly reducing deployment costs. Finally, we explore the feasibility of unifying TTS and ASR tasks using the same data for training, thanks to the simplified pipeline and the S3Tokenizer that reduces the quality requirements for TTS training data.

Xiaomi LLM-Core Team, Zihao Yue, Zhenru Lin et al. · pku

We open-source MiMo-VL-7B-SFT and MiMo-VL-7B-RL, two powerful vision-language models delivering state-of-the-art performance in both general visual understanding and multimodal reasoning. MiMo-VL-7B-RL outperforms Qwen2.5-VL-7B on 35 out of 40 evaluated tasks, and scores 59.4 on OlympiadBench, surpassing models with up to 78B parameters. For GUI grounding applications, it sets a new standard with 56.1 on OSWorld-G, even outperforming specialized models such as UI-TARS. Our training combines four-stage pre-training (2.4 trillion tokens) with Mixed On-policy Reinforcement Learning (MORL) integrating diverse reward signals. We identify the importance of incorporating high-quality reasoning data with long Chain-of-Thought into pre-training stages, and the benefits of mixed RL despite challenges in simultaneous multi-domain optimization. We also contribute a comprehensive evaluation suite covering 50+ tasks to promote reproducibility and advance the field. The model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-VL.

Han Zhu, Wei Kang, Liyong Guo et al.

Generating spoken dialogue is more challenging than monologue text-to-speech (TTS) due to the need for realistic turn-taking and distinct speaker timbres. Existing spoken dialogue generation models, being auto-regressive, suffer from slow and unstable inference. To overcome these limitations, we introduce ZipVoice-Dialog, a non-autoregressive zero-shot spoken dialogue generation model built upon flow matching. Key designs include: 1) speaker-turn embeddings for precise speaker turn-taking; 2) a curriculum learning strategy for stable speech-text alignment; 3) specialized strategies to enable stereo dialogue generation. Additionally, recognizing the lack of open-source large-scale spoken dialogue datasets, we curated OpenDialog, a 6.8k-hour spoken dialogue dataset from in-the-wild speech data. Furthermore, we established a benchmark to comprehensively evaluate various models. Experimental results demonstrate that ZipVoice-Dialog achieves superior performance in intelligibility, speaker turn-taking accuracy, speaker similarity, and inference speed. Our codes, model checkpoints, demo samples, and the OpenDialog dataset are all publicly available at https://github.com/k2-fsa/ZipVoice.

Shengfan Shen, Di Wu, Xingchen Song et al.

Reliable evaluation of modern zero-shot text-to-speech (TTS) models remains challenging. Subjective tests are costly and hard to reproduce, while objective metrics often saturate, failing to distinguish SOTA systems. To address this, we propose Iterate to Differentiate (I2D), an evaluation framework that recursively synthesizes speech using the model's own outputs as references. Higher-quality models exhibit greater resilience to the distributional shift induced by iterative synthesis, resulting in slower performance degradation. I2D exploits this differential degradation to amplify performance gaps and reveal robustness. By aggregating objective metrics across iterations, I2D improves discriminability and alignment with human judgments, increasing system-level SRCC from 0.118 to 0.464 for UTMOSv2. Experiments on 11 models across Chinese, English, and emotion datasets demonstrate that I2D enables more reliable automated evaluation for zero-shot TTS.

Xingchen Song, Di Wu, Dinghao Zhou et al.

Most existing text-to-speech (TTS) systems either synthesize speech sentence by sentence and stitch the results together, or drive synthesis from plain-text dialogues alone. Both approaches leave models with little understanding of global context or paralinguistic cues, making it hard to capture real-world phenomena such as multi-speaker interactions (interruptions, overlapping speech), evolving emotional arcs, and varied acoustic environments. We introduce the Borderless Long Speech Synthesis framework for agent-centric, borderless long audio synthesis. Rather than targeting a single narrow task, the system is designed as a unified capability set spanning VoiceDesigner, multi-speaker synthesis, Instruct TTS, and long-form text synthesis. On the data side, we propose a "Labeling over filtering/cleaning" strategy and design a top-down, multi-level annotation schema we call Global-Sentence-Token. On the model side, we adopt a backbone with a continuous tokenizer and add Chain-of-Thought (CoT) reasoning together with Dimension Dropout, both of which markedly improve instruction following under complex conditions. We further show that the system is Native Agentic by design: the hierarchical annotation doubles as a Structured Semantic Interface between the LLM Agent and the synthesis engine, creating a layered control protocol stack that spans from scene semantics down to phonetic detail. Text thereby becomes an information-complete, wide-band control channel, enabling a front-end LLM to convert inputs of any modality into structured generation commands, extending the paradigm from Text2Speech to borderless long speech synthesis.

Xingchen Song, Di Wu, Binbin Zhang et al.

Scale has opened new frontiers in natural language processing, but at a high cost. In response, by learning to only activate a subset of parameters in training and inference, Mixture-of-Experts (MoE) have been proposed as an energy efficient path to even larger and more capable language models and this shift towards a new generation of foundation models is gaining momentum, particularly within the field of Automatic Speech Recognition (ASR). Recent works that incorporating MoE into ASR models have complex designs such as routing frames via supplementary embedding network, improving multilingual ability for the experts, and utilizing dedicated auxiliary losses for either expert load balancing or specific language handling. We found that delicate designs are not necessary, while an embarrassingly simple substitution of MoE layers for all Feed-Forward Network (FFN) layers is competent for the ASR task. To be more specific, we benchmark our proposed model on a large scale inner-source dataset (160k hours), the results show that we can scale our baseline Conformer (Dense-225M) to its MoE counterparts (MoE-1B) and achieve Dense-1B level Word Error Rate (WER) while maintaining a Dense-225M level Real Time Factor (RTF). Furthermore, by applying Unified 2-pass framework with bidirectional attention decoders (U2++), we achieve the streaming and non-streaming decoding modes in a single MoE based model, which we call U2++ MoE. We hope that our study can facilitate the research on scaling speech foundation models without sacrificing deployment efficiency.

Xingchen Song, Chengdong Liang, Binbin Zhang et al.

Large Automatic Speech Recognition (ASR) models demand a vast number of parameters, copious amounts of data, and significant computational resources during the training process. However, such models can merely be deployed on high-compute cloud platforms and are only capable of performing speech recognition tasks. This leads to high costs and restricted capabilities. In this report, we initially propose the elastic mixture of the expert (eMoE) model. This model can be trained just once and then be elastically scaled in accordance with deployment requirements. Secondly, we devise an unsupervised data creation and validation procedure and gather millions of hours of audio data from diverse domains for training. Using these two techniques, our system achieves elastic deployment capabilities while reducing the Character Error Rate (CER) on the SpeechIO testsets from 4.98\% to 2.45\%. Thirdly, our model is not only competent in Mandarin speech recognition but also proficient in multilingual, multi-dialect, emotion, gender, and sound event perception. We refer to this as Automatic Speech Perception (ASP), and the perception results are presented in the experimental section.

Xingchen Song, Di Wu, Binbin Zhang et al.

In this paper, we present ZeroPrompt (Figure 1-(a)) and the corresponding Prompt-and-Refine strategy (Figure 3), two simple but effective \textbf{training-free} methods to decrease the Token Display Time (TDT) of streaming ASR models \textbf{without any accuracy loss}. The core idea of ZeroPrompt is to append zeroed content to each chunk during inference, which acts like a prompt to encourage the model to predict future tokens even before they were spoken. We argue that streaming acoustic encoders naturally have the modeling ability of Masked Language Models and our experiments demonstrate that ZeroPrompt is engineering cheap and can be applied to streaming acoustic encoders on any dataset without any accuracy loss. Specifically, compared with our baseline models, we achieve 350 $\sim$ 700ms reduction on First Token Display Time (TDT-F) and 100 $\sim$ 400ms reduction on Last Token Display Time (TDT-L), with theoretically and experimentally equal WER on both Aishell-1 and Librispeech datasets.

Xingchen Song, Zhiyong Wu, Yiheng Huang et al.

Non-autoregressive (NAR) transformer models have achieved significantly inference speedup but at the cost of inferior accuracy compared to autoregressive (AR) models in automatic speech recognition (ASR). Most of the NAR transformers take a fixed-length sequence filled with MASK tokens or a redundant sequence copied from encoder states as decoder input, they cannot provide efficient target-side information thus leading to accuracy degradation. To address this problem, we propose a CTC-enhanced NAR transformer, which generates target sequence by refining predictions of the CTC module. Experimental results show that our method outperforms all previous NAR counterparts and achieves 50x faster decoding speed than a strong AR baseline with only 0.0 ~ 0.3 absolute CER degradation on Aishell-1 and Aishell-2 datasets.

Xingchen Song, Guangsen Wang, Zhiyong Wu et al.

Self-attention network (SAN) can benefit significantly from the bi-directional representation learning through unsupervised pretraining paradigms such as BERT and XLNet. In this paper, we present an XLNet-like pretraining scheme "Speech-XLNet" for unsupervised acoustic model pretraining to learn speech representations with SAN. The pretrained SAN is finetuned under the hybrid SAN/HMM framework. We conjecture that by shuffling the speech frame orders, the permutation in Speech-XLNet serves as a strong regularizer to encourage the SAN to make inferences by focusing on global structures through its attention weights. In addition, Speech-XLNet also allows the model to explore the bi-directional contexts for effective speech representation learning. Experiments on TIMIT and WSJ demonstrate that Speech-XLNet greatly improves the SAN/HMM performance in terms of both convergence speed and recognition accuracy compared to the one trained from randomly initialized weights. Our best systems achieve a relative improvement of 11.9% and 8.3% on the TIMIT and WSJ tasks respectively. In particular, the best system achieves a phone error rate (PER) of 13.3% on the TIMIT test set, which to our best knowledge, is the lowest PER obtained from a single system.