Srikanth Madikeri

Iuliia Nigmatulina, Srikanth Madikeri, Esaú Villatoro-Tello et al.

GPU decoding significantly accelerates the output of ASR predictions. While GPUs are already being used for online ASR decoding, post-processing and rescoring on GPUs have not been properly investigated yet. Rescoring with available contextual information can considerably improve ASR predictions. Previous studies have proven the viability of lattice rescoring in decoding and biasing language model (LM) weights in offline and online CPU scenarios. In real-time GPU decoding, partial recognition hypotheses are produced without lattice generation, which makes the implementation of biasing more complex. The paper proposes and describes an approach to integrate contextual biasing in real-time GPU decoding while exploiting the standard Kaldi GPU decoder. Besides the biasing of partial ASR predictions, our approach also permits dynamic context switching allowing a flexible rescoring per each speech segment directly on GPU. The code is publicly released and tested with open-sourced test sets.

Shashi Kumar, Srikanth Madikeri, Juan Zuluaga-Gomez et al.

In traditional conversational intelligence from speech, a cascaded pipeline is used, involving tasks such as voice activity detection, diarization, transcription, and subsequent processing with different NLP models for tasks like semantic endpointing and named entity recognition (NER). Our paper introduces TokenVerse, a single Transducer-based model designed to handle multiple tasks. This is achieved by integrating task-specific tokens into the reference text during ASR model training, streamlining the inference and eliminating the need for separate NLP models. In addition to ASR, we conduct experiments on 3 different tasks: speaker change detection, endpointing, and NER. Our experiments on a public and a private dataset show that the proposed method improves ASR by up to 7.7% in relative WER while outperforming the cascaded pipeline approach in individual task performance. Our code is publicly available: https://github.com/idiap/tokenverse-unifying-speech-nlp

Sergio Burdisso, Séverin Baroudi, Yanis Labrak et al.

We present SDialog, an MIT-licensed open-source Python toolkit that unifies dialog generation, evaluation and mechanistic interpretability into a single end-to-end framework for building and analyzing LLM-based conversational agents. Built around a standardized \texttt{Dialog} representation, SDialog provides: (1) persona-driven multi-agent simulation with composable orchestration for controlled, synthetic dialog generation, (2) comprehensive evaluation combining linguistic metrics, LLM-as-a-judge and functional correctness validators, (3) mechanistic interpretability tools for activation inspection and steering via feature ablation and induction, and (4) audio generation with full acoustic simulation including 3D room modeling and microphone effects. The toolkit integrates with all major LLM backends, enabling mixed-backend experiments under a unified API. By coupling generation, evaluation, and interpretability in a dialog-centric architecture, SDialog enables researchers to build, benchmark and understand conversational systems more systematically.

Sergio Burdisso, Esaú Villatoro-Tello, Srikanth Madikeri et al.

We propose a simple approach for weighting self-connecting edges in a Graph Convolutional Network (GCN) and show its impact on depression detection from transcribed clinical interviews. To this end, we use a GCN for modeling non-consecutive and long-distance semantics to classify the transcriptions into depressed or control subjects. The proposed method aims to mitigate the limiting assumptions of locality and the equal importance of self-connections vs. edges to neighboring nodes in GCNs, while preserving attractive features such as low computational cost, data agnostic, and interpretability capabilities. We perform an exhaustive evaluation in two benchmark datasets. Results show that our approach consistently outperforms the vanilla GCN model as well as previously reported results, achieving an F1=0.84 on both datasets. Finally, a qualitative analysis illustrates the interpretability capabilities of the proposed approach and its alignment with previous findings in psychology.

Esaú Villatoro-Tello, Srikanth Madikeri, Juan Zuluaga-Gomez et al.

In this paper, we perform an exhaustive evaluation of different representations to address the intent classification problem in a Spoken Language Understanding (SLU) setup. We benchmark three types of systems to perform the SLU intent detection task: 1) text-based, 2) lattice-based, and a novel 3) multimodal approach. Our work provides a comprehensive analysis of what could be the achievable performance of different state-of-the-art SLU systems under different circumstances, e.g., automatically- vs. manually-generated transcripts. We evaluate the systems on the publicly available SLURP spoken language resource corpus. Our results indicate that using richer forms of Automatic Speech Recognition (ASR) outputs, namely word-consensus-networks, allows the SLU system to improve in comparison to the 1-best setup (5.5% relative improvement). However, crossmodal approaches, i.e., learning from acoustic and text embeddings, obtains performance similar to the oracle setup, a relative improvement of 17.8% over the 1-best configuration, being a recommended alternative to overcome the limitations of working with automatically generated transcripts.

Iuliia Thorbecke, Esaú Villatoro-Tello, Juan Zuluaga-Gomez et al.

Despite the success of end-to-end automatic speech recognition (ASR) models, challenges persist in recognizing rare, out-of-vocabulary words - including named entities (NE) - and in adapting to new domains using only text data. This work presents a practical approach to address these challenges through an unexplored combination of an NE bias list and a word-level n-gram language model (LM). This solution balances simplicity and effectiveness, improving entities' recognition while maintaining or even enhancing overall ASR performance. We efficiently integrate this enriched biasing method into a transducer-based ASR system, enabling context adaptation with almost no computational overhead. We present our results on three datasets spanning four languages and compare them to state-of-the-art biasing strategies. We demonstrate that the proposed combination of keyword biasing and n-gram LM improves entity recognition by up to 32% relative and reduces overall WER by up to a 12% relative.

Shashi Kumar, Esaú Villatoro-Tello, Sergio Burdisso et al.

Standard LLM-based speech recognition systems typically process utterances in isolation, limiting their ability to leverage conversational context. In this work, we study whether multimodal context from prior turns improves LLM-based ASR and how to represent that context efficiently. We find that, after supervised multi-turn training, conversational context mainly helps with the recognition of contextual entities. However, conditioning on raw context is expensive because the prior-turn audio token sequence grows rapidly with conversation length. To address this, we propose Abstract Compression, which replaces the audio portion of prior turns with a fixed number of learned latent tokens while retaining corresponding transcripts explicitly. On both in-domain and out-of-domain test sets, the compressed model recovers part of the gains of raw-context conditioning with a smaller prior-turn audio footprint. We also provide targeted analyses of the compression setup and its trade-offs.

Rudolf A. Braun, Srikanth Madikeri, Petr Motlicek

A common problem for automatic speech recognition systems is how to recognize words that they did not see during training. Currently there is no established method of evaluating different techniques for tackling this problem. We propose using the CommonVoice dataset to create test sets for multiple languages which have a high out-of-vocabulary (OOV) ratio relative to a training set and release a new tool for calculating relevant performance metrics. We then evaluate, within the context of a hybrid ASR system, how much better subword models are at recognizing OOVs, and how much benefit one can get from incorporating OOV-word information into an existing system by modifying WFSTs. Additionally, we propose a new method for modifying a subword-based language model so as to better recognize OOV-words. We showcase very large improvements in OOV-word recognition and make both the data and code available.

Srikanth Madikeri, Sibo Tong, Juan Zuluaga-Gomez et al.

We present a simple wrapper that is useful to train acoustic models in PyTorch using Kaldi's LF-MMI training framework. The wrapper, called pkwrap (short form of PyTorch kaldi wrapper), enables the user to utilize the flexibility provided by PyTorch in designing model architectures. It exposes the LF-MMI cost function as an autograd function. Other capabilities of Kaldi have also been ported to PyTorch. This includes the parallel training ability when multi-GPU environments are unavailable and decode with graphs created in Kaldi. The package is available on Github at https://github.com/idiap/pkwrap.

Sergio Burdisso, Srikanth Madikeri, Petr Motlicek

Efficiently deriving structured workflows from unannotated dialogs remains an underexplored and formidable challenge in computational linguistics. Automating this process could significantly accelerate the manual design of workflows in new domains and enable the grounding of large language models in domain-specific flowcharts, enhancing transparency and controllability. In this paper, we introduce Dialog2Flow (D2F) embeddings, which differ from conventional sentence embeddings by mapping utterances to a latent space where they are grouped according to their communicative and informative functions (i.e., the actions they represent). D2F allows for modeling dialogs as continuous trajectories in a latent space with distinct action-related regions. By clustering D2F embeddings, the latent space is quantized, and dialogs can be converted into sequences of region/action IDs, facilitating the extraction of the underlying workflow. To pre-train D2F, we build a comprehensive dataset by unifying twenty task-oriented dialog datasets with normalized per-turn action annotations. We also introduce a novel soft contrastive loss that leverages the semantic information of these actions to guide the representation learning process, showing superior performance compared to standard supervised contrastive loss. Evaluation against various sentence embeddings, including dialog-specific ones, demonstrates that D2F yields superior qualitative and quantitative results across diverse domains.

Sergio Burdisso, Esaú Villatoro-Tello, Andrés Carofilis et al.

Adapting automatic speech recognition (ASR) systems based on large language models (LLMs) to new domains using text-only data is a significant yet underexplored challenge. Standard fine-tuning of the LLM on target-domain text often disrupts the critical alignment between speech and text modalities learned by the projector, degrading performance. We introduce a novel text-only adaptation method that emulates the audio projection task by treating it as a text denoising task. Our approach thus trains the LLM to recover clean transcripts from noisy inputs. This process effectively adapts the model to a target domain while preserving cross-modal alignment. Our solution is lightweight, requiring no architectural changes or additional parameters. Extensive evaluation on two datasets demonstrates up to 22.1% relative improvement, outperforming recent state-of-the-art text-only adaptation methods.

Pradeep Rangappa, Andres Carofilis, Jeena Prakash et al.

Fine-tuning pretrained ASR models for specific domains is challenging for small organizations with limited labeled data and computational resources. Here, we explore different data selection pipelines and propose a robust approach that improves ASR adaptation by filtering pseudo-labels generated using Whisper (encoder-decoder) and Zipformer (transducer) models. Our approach integrates multiple selection strategies -- including word error rate (WER) prediction, named entity recognition (NER), and character error rate (CER) analysis -- to extract high-quality training segments. We evaluate our method on Whisper and Zipformer using a 7500-hour baseline, comparing it to a CER-based approach relying on hypotheses from three ASR systems. Fine-tuning on 7500 hours of pseudo-labeled call center data achieves 12.3% WER, while our filtering reduces the dataset to 100 hours (1.4%) with similar performance; a similar trend is observed on Fisher English.

Aref Farhadipour, Teodora Vukovic, Volker Dellwo et al.

Person identification systems often rely on audio, visual, or behavioral cues, but real-world conditions frequently present with missing or degraded modalities. To address this challenge, we propose a multimodal person identification framework incorporating upper-body motion, face, and voice. Experimental results demonstrate that body motion outperforms traditional modalities such as face and voice in within-session evaluations, while serving as a complementary cue that enhances performance in multi-session scenarios. Our model employs a unified hybrid fusion strategy, fusing both feature-level and score-level information to maximize representational richness and decision accuracy. Specifically, it leverages multi-task learning to process modalities independently, followed by cross-attention and gated fusion mechanisms to exploit both unimodal information and cross-modal interactions. Finally, a confidence-weighted strategy and mistake-correction mechanism dynamically adapt to missing data, ensuring that our single classification head achieves optimal performance even in unimodal and bimodal scenarios. We evaluate our method on CANDOR, a newly introduced interview-based multimodal dataset, which we benchmark in this work for the first time. Our results demonstrate that the proposed trimodal system achieves 99.51% Top-1 accuracy on person identification tasks. In addition, we evaluate our model on the VoxCeleb1 dataset as a widely used evaluation protocol and reach 99.92% accuracy in bimodal mode, outperforming conventional approaches. Moreover, we show that our system maintains high accuracy even when one or two modalities are unavailable, making it a robust solution for real-world person recognition applications. The code and data for this work are publicly available.

Sergio Burdisso, Esaú Villatoro-Tello, Shashi Kumar et al.

LLM-based automatic speech recognition (ASR), a well-established approach, connects speech foundation models to large language models (LLMs) through a speech-to-LLM projector, yielding promising results. A common design choice in these architectures is the use of a fixed, manually defined prompt during both training and inference. This setup not only enables applicability across a range of practical scenarios, but also helps maximize model performance. However, the impact of prompt design remains underexplored. This paper presents a comprehensive analysis of commonly used prompts across diverse datasets, showing that prompt choice significantly affects ASR performance and introduces instability, with no single prompt performing best across all cases. Inspired by the speech-to-LLM projector, we propose a prompt projector module, a simple, model-agnostic extension that learns to project prompt embeddings to more effective regions of the LLM input space, without modifying the underlying LLM-based ASR model. Experiments on four datasets show that the addition of a prompt projector consistently improves performance, reduces variability, and outperforms the best manually selected prompts.

Shashi Kumar, Srikanth Madikeri, Esaú Villatoro-Tello et al.

Token-based multitasking frameworks like TokenVerse require all training utterances to have labels for all tasks, hindering their ability to leverage partially annotated datasets and scale effectively. We propose TokenVerse++, which introduces learnable vectors in the acoustic embedding space of the XLSR-Transducer ASR model for dynamic task activation. This core mechanism enables training with utterances labeled for only a subset of tasks, a key advantage over TokenVerse. We demonstrate this by successfully integrating a dataset with partial labels, specifically for ASR and an additional task, language identification, improving overall performance. TokenVerse++ achieves results on par with or exceeding TokenVerse across multiple tasks, establishing it as a more practical multitask alternative without sacrificing ASR performance.

Andres Carofilis, Pradeep Rangappa, Srikanth Madikeri et al.

Fine-tuning pretrained ASR models for specific domains is challenging when labeled data is scarce. But unlabeled audio and labeled data from related domains are often available. We propose an incremental semi-supervised learning pipeline that first integrates a small in-domain labeled set and an auxiliary dataset from a closely related domain, achieving a relative improvement of 4% over no auxiliary data. Filtering based on multi-model consensus or named entity recognition (NER) is then applied to select and iteratively refine pseudo-labels, showing slower performance saturation compared to random selection. Evaluated on the multi-domain Wow call center and Fisher English corpora, it outperforms single-step fine-tuning. Consensus-based filtering outperforms other methods, providing up to 22.3% relative improvement on Wow and 24.8% on Fisher over single-step fine-tuning with random selection. NER is the second-best filter, providing competitive performance at a lower computational cost.

Juan Zuluaga-Gomez, Iuliia Nigmatulina, Amrutha Prasad et al.

Voice communication between air traffic controllers (ATCos) and pilots is critical for ensuring safe and efficient air traffic control (ATC). This task requires high levels of awareness from ATCos and can be tedious and error-prone. Recent attempts have been made to integrate artificial intelligence (AI) into ATC in order to reduce the workload of ATCos. However, the development of data-driven AI systems for ATC demands large-scale annotated datasets, which are currently lacking in the field. This paper explores the lessons learned from the ATCO2 project, a project that aimed to develop a unique platform to collect and preprocess large amounts of ATC data from airspace in real time. Audio and surveillance data were collected from publicly accessible radio frequency channels with VHF receivers owned by a community of volunteers and later uploaded to Opensky Network servers, which can be considered an "unlimited source" of data. In addition, this paper reviews previous work from ATCO2 partners, including (i) robust automatic speech recognition, (ii) natural language processing, (iii) English language identification of ATC communications, and (iv) the integration of surveillance data such as ADS-B. We believe that the pipeline developed during the ATCO2 project, along with the open-sourcing of its data, will encourage research in the ATC field. A sample of the ATCO2 corpus is available on the following website: https://www.atco2.org/data, while the full corpus can be purchased through ELDA at http://catalog.elra.info/en-us/repository/browse/ELRA-S0484. We demonstrated that ATCO2 is an appropriate dataset to develop ASR engines when little or near to no ATC in-domain data is available. For instance, with the CNN-TDNNf kaldi model, we reached the performance of as low as 17.9% and 24.9% WER on public ATC datasets which is 6.6/7.6% better than "out-of-domain" but supervised CNN-TDNNf model.

Apoorv Vyas, Srikanth Madikeri, Hervé Bourlard

In this work, we investigate if the wav2vec 2.0 self-supervised pretraining helps mitigate the overfitting issues with connectionist temporal classification (CTC) training to reduce its performance gap with flat-start lattice-free MMI (E2E-LFMMI) for automatic speech recognition with limited training data. Towards that objective, we use the pretrained wav2vec 2.0 BASE model and fine-tune it on three different datasets including out-of-domain (Switchboard) and cross-lingual (Babel) scenarios. Our results show that for supervised adaptation of the wav2vec 2.0 model, both E2E-LFMMI and CTC achieve similar results; significantly outperforming the baselines trained only with supervised data. Fine-tuning the wav2vec 2.0 model with E2E-LFMMI and CTC we obtain the following relative WER improvements over the supervised baseline trained with E2E-LFMMI. We get relative improvements of 40% and 44% on the clean-set and 64% and 58% on the test set of Librispeech (100h) respectively. On Switchboard (300h) we obtain relative improvements of 33% and 35% respectively. Finally, for Babel languages, we obtain relative improvements of 26% and 23% on Swahili (38h) and 18% and 17% on Tagalog (84h) respectively.

Apoorv Vyas, Srikanth Madikeri, Hervé Bourlard

In this work, we propose lattice-free MMI (LFMMI) for supervised adaptation of self-supervised pretrained acoustic model. We pretrain a Transformer model on thousand hours of untranscribed Librispeech data followed by supervised adaptation with LFMMI on three different datasets. Our results show that fine-tuning with LFMMI, we consistently obtain relative WER improvements of 10% and 35.3% on the clean and other test sets of Librispeech (100h), 10.8% on Switchboard (300h), and 4.3% on Swahili (38h) and 4.4% on Tagalog (84h) compared to the baseline trained only with supervised data.

Seyyed Saeed Sarfjoo, Srikanth Madikeri, Petr Motlicek

To better model the contextual information and increase the generalization ability of Speech Activity Detection (SAD) system, this paper leverages a multi-lingual Automatic Speech Recognition (ASR) system to perform SAD. Sequence discriminative training of Acoustic Model (AM) using Lattice-Free Maximum Mutual Information (LF-MMI) loss function, effectively extracts the contextual information of the input acoustic frame. Multi-lingual AM training, causes the robustness to noise and language variabilities. The index of maximum output posterior is considered as a frame-level speech/non-speech decision function. Majority voting and logistic regression are applied to fuse the language-dependent decisions. The multi-lingual ASR is trained on 18 languages of BABEL datasets and the built SAD is evaluated on 3 different languages. On out-of-domain datasets, the proposed SAD model shows significantly better performance with respect to baseline models. On the Ester2 dataset, without using any in-domain data, this model outperforms the WebRTC, phoneme recognizer based VAD (Phn Rec), and Pyannote baselines (respectively by 7.1, 1.7, and 2.7% absolute) in Detection Error Rate (DetER) metrics. Similarly, on the LiveATC dataset, this model outperforms the WebRTC, Phn Rec, and Pyannote baselines (respectively by 6.4, 10.0, and 3.7% absolutely) in DetER metrics.

Amrutha Prasad, Petr Motlicek, Srikanth Madikeri

State-of-the-art hybrid automatic speech recognition (ASR) system exploits deep neural network (DNN) based acoustic models (AM) trained with Lattice Free-Maximum Mutual Information (LF-MMI) criterion and n-gram language models. The AMs typically have millions of parameters and require significant parameter reduction to operate on embedded devices. The impact of parameter quantization on the overall word recognition performance is studied in this paper. Following approaches are presented: (i) AM trained in Kaldi framework with conventional factorized TDNN (TDNN-F) architecture, (ii) the TDNN AM built in Kaldi loaded into the PyTorch toolkit using a C++ wrapper for post-training quantization, (iii) quantization-aware training in PyTorch for Kaldi TDNN model, (iv) quantization-aware training in Kaldi. Results obtained on standard Librispeech setup provide an interesting overview of recognition accuracy w.r.t. applied quantization scheme.

Mael Fabien, Seyyed Saeed Sarfjoo, Petr Motlicek et al.

Criminal investigations mostly rely on the collection of speech conversational data in order to identify speakers and build or enrich an existing criminal network. Social network analysis tools are then applied to identify the most central characters and the different communities within the network. We introduce two candidate datasets for criminal conversational data, Crime Scene Investigation (CSI), a television show, and the ROXANNE simulated data. We also introduce the metric of conversation accuracy in the context of criminal investigations. By re-ranking candidate speakers based on the frequency of previous interactions, we improve the speaker identification baseline by 1.2% absolute (1.3% relative), and the conversation accuracy by 2.6% absolute (3.4% relative) on CSI data, and by 1.1% absolute (1.2% relative), and 2% absolute (2.5% relative) respectively on the ROXANNE simulated data.

Navid Rekabsaz, Nikolaos Pappas, James Henderson et al.

Neural language modeling (LM) has led to significant improvements in several applications, including Automatic Speech Recognition. However, they typically require large amounts of training data, which is not available for many domains and languages. In this study, we propose a multilingual neural language model architecture, trained jointly on the domain-specific data of several low-resource languages. The proposed multilingual LM consists of language specific word embeddings in the encoder and decoder, and one language specific LSTM layer, plus two LSTM layers with shared parameters across the languages. This multilingual LM model facilitates transfer learning across the languages, acting as an extra regularizer in very low-resource scenarios. We integrate our proposed multilingual approach with a state-of-the-art highly-regularized neural LM, and evaluate on the conversational data domain for four languages over a range of training data sizes. Compared to monolingual LMs, the results show significant improvements of our proposed multilingual LM when the amount of available training data is limited, indicating the advantages of cross-lingual parameter sharing in very low-resource language modeling.

Nauman Dawalatabad, Srikanth Madikeri, C Chandra Sekhar et al.

The two-pass information bottleneck (TPIB) based speaker diarization system operates independently on different conversational recordings. TPIB system does not consider previously learned speaker discriminative information while diarizing new conversations. Hence, the real time factor (RTF) of TPIB system is high owing to the training time required for the artificial neural network (ANN). This paper attempts to improve the RTF of the TPIB system using an incremental transfer learning approach where the parameters learned by the ANN from other conversations are updated using current conversation rather than learning parameters from scratch. This reduces the RTF significantly. The effectiveness of the proposed approach compared to the baseline IB and the TPIB systems is demonstrated on standard NIST and AMI conversational meeting datasets. With a minor degradation in performance, the proposed system shows a significant improvement of 33.07% and 24.45% in RTF with respect to TPIB system on the NIST RT-04Eval and AMI-1 datasets, respectively.