TokenSplit: Using Discrete Speech Representations for Direct, Refined, and Transcript-Conditioned Speech Separation and Recognition

We present TokenSplit, a speech separation model that acts on discrete token sequences. The model is trained on multiple tasks simultaneously: separate and transcribe each speech source, and generate speech from text. The model operates on transcripts and audio token sequences and achieves multiple tasks through masking of inputs. The model is a sequence-to-sequence encoder-decoder model that uses the Transformer architecture. We also present a "refinement" version of the model that predicts enhanced audio tokens from the audio tokens of speech separated by a conventional separation model. Using both objective metrics and subjective MUSHRA listening tests, we show that our model achieves excellent performance in terms of separation, both with or without transcript conditioning. We also measure the automatic speech recognition (ASR) performance and provide audio samples of speech synthesis to demonstrate the additional utility of our model.

UniverSLU: Universal Spoken Language Understanding for Diverse Classification and Sequence Generation Tasks with a Single Network

Recent studies have demonstrated promising outcomes by employing large language models with multi-tasking capabilities. They utilize prompts to guide the model's behavior and surpass performance of task-specific models. Motivated by this, we ask: can we build a single model that jointly perform various spoken language understanding (SLU) tasks? To address this, we utilize pre-trained automatic speech recognition (ASR) models and employ various task and dataset specifiers as discrete prompts. We demonstrate efficacy of our single multi-task learning (MTL) model "UniverSLU" for 12 different speech classification and sequence generation tasks across 17 datasets and 9 languages. Results show that UniverSLU achieves competitive performance and even surpasses task-specific models. We also conduct preliminary investigations into enabling human-interpretable natural phrases instead of task specifiers as discrete prompts and test the model's generalization capabilities to new paraphrases.

Boosting End-to-End Multilingual Phoneme Recognition through Exploiting Universal Speech Attributes Constraints

We propose a first step toward multilingual end-to-end automatic speech recognition (ASR) by integrating knowledge about speech articulators. The key idea is to leverage a rich set of fundamental units that can be defined "universally" across all spoken languages, referred to as speech attributes, namely manner and place of articulation. Specifically, several deterministic attribute-to-phoneme mapping matrices are constructed based on the predefined set of universal attribute inventory, which projects the knowledge-rich articulatory attribute logits, into output phoneme logits. The mapping puts knowledge-based constraints to limit inconsistency with acoustic-phonetic evidence in the integrated prediction. Combined with phoneme recognition, our phone recognizer is able to infer from both attribute and phoneme information. The proposed joint multilingual model is evaluated through phoneme recognition. In multilingual experiments over 6 languages on benchmark datasets LibriSpeech and CommonVoice, we find that our proposed solution outperforms conventional multilingual approaches with a relative improvement of 6.85% on average, and it also demonstrates a much better performance compared to monolingual model. Further analysis conclusively demonstrates that the proposed solution eliminates phoneme predictions that are inconsistent with attributes.

SALMONN: Towards Generic Hearing Abilities for Large Language Models

Hearing is arguably an essential ability of artificial intelligence (AI) agents in the physical world, which refers to the perception and understanding of general auditory information consisting of at least three types of sounds: speech, audio events, and music. In this paper, we propose SALMONN, a speech audio language music open neural network, built by integrating a pre-trained text-based large language model (LLM) with speech and audio encoders into a single multimodal model. SALMONN enables the LLM to directly process and understand general audio inputs and achieve competitive performances on a number of speech and audio tasks used in training, such as automatic speech recognition and translation, auditory-information-based question answering, emotion recognition, speaker verification, and music and audio captioning \textit{etc.} SALMONN also has a diverse set of emergent abilities unseen in the training, which includes but is not limited to speech translation to untrained languages, speech-based slot filling, spoken-query-based question answering, audio-based storytelling, and speech audio co-reasoning \textit{etc}. The presence of the cross-modal emergent abilities is studied, and a novel few-shot activation tuning approach is proposed to activate such abilities of SALMONN. To our knowledge, SALMONN is the first model of its type and can be regarded as a step towards AI with generic hearing abilities. An interactive demo of SALMONN is available at \texttt{\url{https://github.com/bytedance/SALMONN}}, and the training code and model checkpoints will be released upon acceptance.

Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition

Conformer-based models have become the most dominant end-to-end architecture for speech processing tasks. In this work, we propose a carefully redesigned Conformer with a new down-sampling schema. The proposed model, named Fast Conformer, is 2.8x faster than original Conformer, while preserving state-of-the-art accuracy on Automatic Speech Recognition benchmarks. Also we replace the original Conformer global attention with limited context attention post-training to enable transcription of an hour-long audio. We further improve long-form speech transcription by adding a global token. Fast Conformer combined with a Transformer decoder also outperforms the original Conformer in accuracy and in speed for Speech Translation and Spoken Language Understanding.

LAE-ST-MoE: Boosted Language-Aware Encoder Using Speech Translation Auxiliary Task for E2E Code-switching ASR

Recently, to mitigate the confusion between different languages in code-switching (CS) automatic speech recognition (ASR), the conditionally factorized models, such as the language-aware encoder (LAE), explicitly disregard the contextual information between different languages. However, this information may be helpful for ASR modeling. To alleviate this issue, we propose the LAE-ST-MoE framework. It incorporates speech translation (ST) tasks into LAE and utilizes ST to learn the contextual information between different languages. It introduces a task-based mixture of expert modules, employing separate feed-forward networks for the ASR and ST tasks. Experimental results on the ASRU 2019 Mandarin-English CS challenge dataset demonstrate that, compared to the LAE-based CTC, the LAE-ST-MoE model achieves a 9.26% mix error reduction on the CS test with the same decoding parameter. Moreover, the well-trained LAE-ST-MoE model can perform ST tasks from CS speech to Mandarin or English text.

Adaptive Contextual Biasing for Transducer Based Streaming Speech Recognition

By incorporating additional contextual information, deep biasing methods have emerged as a promising solution for speech recognition of personalized words. However, for real-world voice assistants, always biasing on such personalized words with high prediction scores can significantly degrade the performance of recognizing common words. To address this issue, we propose an adaptive contextual biasing method based on Context-Aware Transformer Transducer (CATT) that utilizes the biased encoder and predictor embeddings to perform streaming prediction of contextual phrase occurrences. Such prediction is then used to dynamically switch the bias list on and off, enabling the model to adapt to both personalized and common scenarios. Experiments on Librispeech and internal voice assistant datasets show that our approach can achieve up to 6.7% and 20.7% relative reduction in WER and CER compared to the baseline respectively, mitigating up to 96.7% and 84.9% of the relative WER and CER increase for common cases. Furthermore, our approach has a minimal performance impact in personalized scenarios while maintaining a streaming inference pipeline with negligible RTF increase.

MuAViC: A Multilingual Audio-Visual Corpus for Robust Speech Recognition and Robust Speech-to-Text Translation

We introduce MuAViC, a multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation providing 1200 hours of audio-visual speech in 9 languages. It is fully transcribed and covers 6 English-to-X translation as well as 6 X-to-English translation directions. To the best of our knowledge, this is the first open benchmark for audio-visual speech-to-text translation and the largest open benchmark for multilingual audio-visual speech recognition. Our baseline results show that MuAViC is effective for building noise-robust speech recognition and translation models. We make the corpus available at https://github.com/facebookresearch/muavic.

Multi-View Frequency-Attention Alternative to CNN Frontends for Automatic Speech Recognition

Convolutional frontends are a typical choice for Transformer-based automatic speech recognition to preprocess the spectrogram, reduce its sequence length, and combine local information in time and frequency similarly. However, the width and height of an audio spectrogram denote different information, e.g., due to reverberation as well as the articulatory system, the time axis has a clear left-to-right dependency. On the contrary, vowels and consonants demonstrate very different patterns and occupy almost disjoint frequency ranges. Therefore, we hypothesize, global attention over frequencies is beneficial over local convolution. We obtain 2.4 % relative word error rate reduction (rWERR) on a production scale Conformer transducer replacing its convolutional neural network frontend by the proposed F-Attention module on Alexa traffic. To demonstrate generalizability, we validate this on public LibriSpeech data with a long short term memory-based listen attend and spell architecture obtaining 4.6 % rWERR and demonstrate robustness to (simulated) noisy conditions.

The Gift of Feedback: Improving ASR Model Quality by Learning from User Corrections through Federated Learning

Automatic speech recognition (ASR) models are typically trained on large datasets of transcribed speech. As language evolves and new terms come into use, these models can become outdated and stale. In the context of models trained on the server but deployed on edge devices, errors may result from the mismatch between server training data and actual on-device usage. In this work, we seek to continually learn from on-device user corrections through Federated Learning (FL) to address this issue. We explore techniques to target fresh terms that the model has not previously encountered, learn long-tail words, and mitigate catastrophic forgetting. In experimental evaluations, we find that the proposed techniques improve model recognition of fresh terms, while preserving quality on the overall language distribution.