We propose self-speculative decoding for speech-aware LLMs by using the CTC encoder as a draft model to accelerate auto-regressive (AR) inference and improve ASR accuracy. Our three-step procedure works as follows: (1) if the frame entropies of the CTC output distributions are below a threshold, the greedy CTC hypothesis is accepted as final; (2) otherwise, the CTC hypothesis is verified in a single LLM forward pass using a relaxed acceptance criterion based on token likelihoods; (3) if verification fails, AR decoding resumes from the accepted CTC prefix. Experiments on nine corpora and five languages show that this approach can simultaneously accelerate decoding and reduce WER. On the HuggingFace Open ASR benchmark with a 1B parameter LLM and 440M parameter CTC encoder, we achieve a record 5.58

Speaker-Attributed Automatic Speech Recognition (SAA) enhances traditional ASR systems by incorporating relative speaker identity tags directly into the transcript (e.g., [Speaker 1]:, [Speaker 2]:). In this work, we extend the capabilities of Granite-speech, a state-of-the-art speech-aware Large Language Model (LLM) originally trained for transcription and translation. We demonstrate that it can be effectively adapted for SAA with only minimal architectural changes. Our core contribution is the introduction of speaker cluster identification tags (e.g., [Speaker 1 cluster 42]:) which are jointly trained with SAA to significantly improve accuracy. To address limitations in training data, we propose a data augmentation method that uses artificially concatenated multi-speaker conversations. Our approach is evaluated across multiple benchmarks and shows superior performance compared to conventional pipelines that sequentially perform speaker diarization followed by ASR.

While autoregressive (AR) LLM-based ASR systems achieve strong accuracy, their sequential decoding limits parallelism and incurs high latency. We propose NLE, a non-autoregressive (NAR) approach that formulates speech recognition as conditional transcript editing, enabling fully parallel prediction. NLE extracts acoustic embeddings and an initial hypothesis from a pretrained speech encoder, then refines the hypothesis using a bidirectional LLM editor trained with a latent alignment objective. An interleaved padding strategy exploits the identity mapping bias of Transformers, allowing the model to focus on corrections rather than full reconstruction. On the Open ASR leaderboard, NLE++ achieves 5.67% average WER with an RTFx (inverse real-time factor) of 1630. In single-utterance scenarios, NLE achieves 27x speedup over the AR baseline, making it suitable for real-time applications.

Speech-aware LLMs (SLLMs) have recently achieved state-of-the-art ASR performance; however, they still fail to accurately transcribe bias words that appear rarely or never in the training data. Contextual biasing mechanisms are commonly implemented by introducing a predefined bias word list into the model via a text prompt or additional module. For further improvement, predefined bias words can be paired with their phoneme representations as pronunciation cues. Typically, phoneme sequences are generated through a G2P system that covers the target languages and domains of the bias words. Therefore, when a compatible G2P system is unavailable, phoneme-assisted contextual biasing becomes difficult to perform. Moreover, manually adding accurate phoneme sequences requires advanced phonetic knowledge. In this paper, we explore contextual biasing in SLLM based on acoustic cues associated with a set of common words whose pronunciations are partially similar to those of the target bias words. We assume ASR applications in which end users do not require special knowledge of phonetics or utilize G2P tools for inference. For enhanced robustness, we also introduce bias word positional prediction implemented in a multi-output learning fashion. Our method reduces bias word recognition errors by 16.3

Recent advances in speech-aware language models have coupled strong acoustic encoders with large language models, enabling systems that move beyond transcription to produce richer outputs. Among these, word-level timestamp prediction is critical for applications such as captioning, media search, and multimodal synchronization, yet it is often handled by external alignment tools. In this work, we extend an existing speech-aware language model to predict timestamps directly alongside transcripts. We introduce a set of novel lightweight training strategies that improve alignment robustness while preserving recognition quality. Experiments across multiple datasets show that these strategies not only enhance timestamp accuracy, but also yield gains in overall ASR performance. Together, they demonstrate an efficient and unified approach to speech recognition with precise timestamp prediction.

In this paper, we take a step towards jointly modeling automatic speech recognition (STT) and speech synthesis (TTS) in a fully non-autoregressive way. We develop a novel multimodal framework capable of handling the speech and text modalities as input either individually or together. The proposed model can also be trained with unpaired speech or text data owing to its multimodal nature. We further propose an iterative refinement strategy to improve the STT and TTS performance of our model such that the partial hypothesis at the output can be fed back to the input of our model, thus iteratively improving both STT and TTS predictions. We show that our joint model can effectively perform both STT and TTS tasks, outperforming the STT-specific baseline in all tasks and performing competitively with the TTS-specific baseline across a wide range of evaluation metrics.

There is an on-going body of research on training separate dedicated models for either short-form or long-form utterances. Multi-form acoustic models that are simply trained on combined data from long-form and short-form utterances often suffer from various negative impacts due to the diversity of a speaking style, an accent, and a recording condition. In addition, a linguistic mismatch that comes from an utterance length is also another factor of the degradation. In this paper we investigate novel techniques for training unified Conformer-based models on multi-form speech data obtained from diverse domains and sources to serve multiple downstream applications with a single model. Our approach incorporates chunk-wise short-term discriminative knowledge distillation with an encoder embedding masking and mitigates the aforementioned problems that appear for single unified models. We show the benefit of our proposed technique on long and short-form ASR test sets by comparing our models against several variants trained by mixing utterances with various audio lengths. The proposed technique provides a significant improvement of up to 8.5% relative WER reduction over baseline systems that operate at a similar decoding cost.

Granite-speech LLMs are compact and efficient speech language models specifically designed for English ASR and automatic speech translation (AST). The models were trained by modality aligning the 2B and 8B parameter variants of granite-3.3-instruct to speech on publicly available open-source corpora containing audio inputs and text targets consisting of either human transcripts for ASR or automatically generated translations for AST. Comprehensive benchmarking shows that on English ASR, which was our primary focus, they outperform several competitors' models that were trained on orders of magnitude more proprietary data, and they keep pace on English-to-X AST for major European languages, Japanese, and Chinese. The speech-specific components are: a conformer acoustic encoder using block attention and self-conditioning trained with connectionist temporal classification, a windowed query-transformer speech modality adapter used to do temporal downsampling of the acoustic embeddings and map them to the LLM text embedding space, and LoRA adapters to further fine-tune the text LLM. Granite-speech-3.3 operates in two modes: in speech mode, it performs ASR and AST by activating the encoder, projector, and LoRA adapters; in text mode, it calls the underlying granite-3.3-instruct model directly (without LoRA), essentially preserving all the text LLM capabilities and safety. Both models are freely available on HuggingFace (https://huggingface.co/ibm-granite/granite-speech-3.3-2b and https://huggingface.co/ibm-granite/granite-speech-3.3-8b) and can be used for both research and commercial purposes under a permissive Apache 2.0 license.

Conformer CTC-Encoders have consistently delivered state-of-the-art results in the field of Automatic Speech Recognition (ASR); however, their merits for tasks that demand more semantic and paralinguistic information, such as Automatic Speech Understanding (ASU), Speech Emotion Recognition (SER) and Speech Translation (ST), still need further investigation. In this paper, we introduce a Speech Large Language Model (SLLM) system based on a Conformer CTC-Encoder and on the Granite Large Language Model that allowed us to perform several experiments on ASR, SER and ST tasks. These experiments have not only confirmed the strength of Conformer CTC-encoders for ASR, but also, they have shown that the outputs of intermediate Conformer Blocks, of the Conformer CTC-Encoder, carry important information for SER tasks and that the Conformer CTC-Encoder can be efficiently fine-tuned for SER tasks.

Limited transcribed spoken style data is a critical bottleneck in building automatic speech recognition (ASR) systems for low-resource languages. Prompting a large language model (LLM) to paraphrase input text can generate novel text data that is constrained to be semantically similar to the source data. We leverage this capability of LLMs to improve the performance of low-resource ASR systems by increasing the limited text training data while keeping the same spoken style. Since word sequences in the training data are now more diverse and the vocabulary of the ASR model is also expanded, this approach allows for building general purpose ASR without prior knowledge of various domains in the low-resource language. In our experiments with Brazilian Portuguese as a low-resource language, paraphrased data enhanced the n-gram language model (LM) used to build the weighted finite state transducer (WFST) for decoding with a Conformer-CTC speech recognition model, resulting in improvement of word error rate (WER) by 15.6% over the baseline model. Synthesizing the paraphrased text into speech and using it to fine-tune the acoustic model (AM) component helped to further improve the WER by 2.9%, achieving a combined improvement of 18.5%. We also demonstrate the usefulness of our proposed approach for high-resource languages like English.

The emergence of industrial-scale automatic speech recognition (ASR) models such as Whisper and USM, trained on 1M hours of weakly labelled and 12M hours of audio only proprietary data respectively, has led to a stronger need for large scale public ASR corpora and competitive open source pipelines. Unlike the said models, large language models are typically based on Transformer decoders, and it remains unclear if decoder-only models trained on public data alone can deliver competitive performance. In this work, we investigate factors such as choice of training datasets and modeling components necessary for obtaining the best performance using only public English ASR corpora. Our Decoder-Only Transformer for ASR (DOTA) model comprehensively outperforms the encoder-decoder open source replication of Whisper (OWSM) on nearly all English ASR benchmarks and outperforms Whisper large-v3 on 6 out of 15 test sets. We release our codebase and model checkpoints under permissive license.

Transferring the knowledge of large language models (LLMs) is a promising technique to incorporate linguistic knowledge into end-to-end automatic speech recognition (ASR) systems. However, existing works only transfer a single representation of LLM (e.g. the last layer of pretrained BERT), while the representation of a text is inherently non-unique and can be obtained variously from different layers, contexts and models. In this work, we explore a wide range of techniques to obtain and transfer multiple representations of LLMs into a transducer-based ASR system. While being conceptually simple, we show that transferring multiple representations of LLMs can be an effective alternative to transferring only a single LLM representation.

Non-autoregressive (NAR) modeling has gained significant interest in speech processing since these models achieve dramatically lower inference time than autoregressive (AR) models while also achieving good transcription accuracy. Since NAR automatic speech recognition (ASR) models must wait for the completion of the entire utterance before processing, some works explore streaming NAR models based on blockwise attention for low-latency applications. However, streaming NAR models significantly lag in accuracy compared to streaming AR and non-streaming NAR models. To address this, we propose a streaming "semi-autoregressive" ASR model that incorporates the labels emitted in previous blocks as additional context using a Language Model (LM) subnetwork. We also introduce a novel greedy decoding algorithm that addresses insertion and deletion errors near block boundaries while not significantly increasing the inference time. Experiments show that our method outperforms the existing streaming NAR model by 19% relative on Tedlium2, 16%/8% on Librispeech-100 clean/other test sets, and 19%/8% on the Switchboard(SWB)/Callhome(CH) test sets. It also reduced the accuracy gap with streaming AR and non-streaming NAR models while achieving 2.5x lower latency. We also demonstrate that our approach can effectively utilize external text data to pre-train the LM subnetwork to further improve streaming ASR accuracy.

We improve on the popular conformer architecture by replacing the depthwise temporal convolutions with diagonal state space (DSS) models. DSS is a recently introduced variant of linear RNNs obtained by discretizing a linear dynamical system with a diagonal state transition matrix. DSS layers project the input sequence onto a space of orthogonal polynomials where the choice of basis functions, metric and support is controlled by the eigenvalues of the transition matrix. We compare neural transducers with either conformer or our proposed DSS-augmented transformer (DSSformer) encoders on three public corpora: Switchboard English conversational telephone speech 300 hours, Switchboard+Fisher 2000 hours, and a spoken archive of holocaust survivor testimonials called MALACH 176 hours. On Switchboard 300/2000 hours, we reach a single model performance of 8.9%/6.7% WER on the combined test set of the Hub5 2000 evaluation, respectively, and on MALACH we improve the WER by 7% relative over the previous best published result. In addition, we present empirical evidence suggesting that DSS layers learn damped Fourier basis functions where the attenuation coefficients are layer specific whereas the frequency coefficients converge to almost identical linearly-spaced values across all layers.

In this paper we investigate several techniques for improving the performance of RNN transducer (RNNT) acoustic models for conversational speech recognition and report state-of-the-art word error rates (WERs) on the 2000-hour Switchboard dataset. We show that n-best label smoothing and length perturbation which show improved performance on the smaller 300-hour dataset are also very effective on large datasets. We further give a rigorous theoretical interpretation of the n-best label smoothing based on stochastic approximation for training RNNT under the maximum likelihood criterion. Random quantization is also introduced to improve the generalization of RNNT models. On the 2000-hour Switchboard dataset, we report a single model performance of 4.9% and 7.7% WERs on the Switchboard and CallHome portions of NIST Hub5 2000, 7.1% on NIST Hub5 2001 and 6.8% on NIST RT03, without using external LMs.

Soft random sampling (SRS) is a simple yet effective approach for efficient training of large-scale deep neural networks when dealing with massive data. SRS selects a subset uniformly at random with replacement from the full data set in each epoch. In this paper, we conduct a theoretical and empirical analysis of SRS. First, we analyze its sampling dynamics including data coverage and occupancy. Next, we investigate its convergence with non-convex objective functions and give the convergence rate. Finally, we provide its generalization performance. We empirically evaluate SRS for image recognition on CIFAR10 and automatic speech recognition on Librispeech and an in-house payload dataset to demonstrate its effectiveness. Compared to existing coreset-based data selection methods, SRS offers a better accuracy-efficiency trade-off. Especially on real-world industrial scale data sets, it is shown to be a powerful training strategy with significant speedup and competitive performance with almost no additional computing cost.

Publicly available datasets traditionally used to train E2E ASR models for conversational telephone speech recognition are based on clean, short duration, single speaker utterances collected on separate channels. While E2E ASR models achieve state-of-the-art performance on recognition tasks that match well with such training data, they are observed to fail on test recordings that contain multiple speakers, significant channel or background noise or span longer durations than training data utterances. To mitigate these issues, we propose an on-the-fly data augmentation strategy that transforms single speaker training data into multiple speaker data by appending together multiple single speaker utterances. The proposed technique encourages the E2E model to become robust to speaker changes and also process longer utterances effectively. During training, the model is also guided by a teacher model trained on single speaker utterances to map its multi-speaker encoder embeddings to better performing single speaker representations. With the proposed technique we obtain 7-14% relative improvement on various single speaker and multiple speaker test sets. We also show how this technique is able to improve recognition performance by up to 14% by capturing useful information from preceding spoken utterances used as dialog history.

Despite the impressive performance of ASR models on mainstream benchmarks, their performance on rare words is unsatisfactory. In enterprise settings, often a focused list of entities (such as locations, names, etc) are available which can be used to adapt the model to the terminology of specific domains. In this paper, we present a novel inference algorithm that improves the prediction of state-of-the-art ASR models using nearest-neighbor-based matching on an inference-time word list. We consider both the Transducer architecture that is useful in the streaming setting, and state-of-the-art encoder-decoder models such as Whisper. In our approach, a list of rare entities is indexed in a memory by synthesizing speech for each entry, and then storing the internal acoustic and language model states obtained from the best possible alignment on the ASR model. The memory is organized as a trie which we harness to perform a stateful lookup during inference. A key property of our extension is that we prevent spurious matches by restricting to only word-level matches. In our experiments on publicly available datasets and private benchmarks, we show that our method is effective in significantly improving rare word recognition.

We investigate soft sampling which is a simple yet effective approach for efficient training of large-scale deep neural network models when dealing with massive data. Soft sampling selects a subset uniformly at random with replacement from the full data set in each epoch. First, we derive a theoretical convergence guarantee for soft sampling on non-convex objective functions and give the convergence rate. Next, we analyze the data coverage and occupancy properties of soft sampling from the perspective of the coupon collector's problem. And finally, we evaluate soft sampling on various machine learning tasks using various network architectures and demonstrate its effectiveness. Compared to existing coreset-based data selection methods, soft sampling offers a better accuracy-efficiency trade-off. Especially on real-world industrial scale data sets, soft sampling can achieve significant speedup and competitive performance with almost no additional computing cost.

We report on aggressive quantization strategies that greatly accelerate inference of Recurrent Neural Network Transducers (RNN-T). We use a 4 bit integer representation for both weights and activations and apply Quantization Aware Training (QAT) to retrain the full model (acoustic encoder and language model) and achieve near-iso-accuracy. We show that customized quantization schemes that are tailored to the local properties of the network are essential to achieve good performance while limiting the computational overhead of QAT. Density ratio Language Model fusion has shown remarkable accuracy gains on RNN-T workloads but it severely increases the computational cost of inference. We show that our quantization strategies enable using large beam widths for hypothesis search while achieving streaming-compatible runtimes and a full model compression ratio of 7.6$\times$ compared to the full precision model. Via hardware simulations, we estimate a 3.4$\times$ acceleration from FP16 to INT4 for the end-to-end quantized RNN-T inclusive of LM fusion, resulting in a Real Time Factor (RTF) of 0.06. On the NIST Hub5 2000, Hub5 2001, and RT-03 test sets, we retain most of the gains associated with LM fusion, improving the average WER by $>$1.5%.