Mikko Kurimo

Anssi Moisio, Dejan Porjazovski, Aku Rouhe et al.

The Donate Speech campaign has so far succeeded in gathering approximately 3600 hours of ordinary, colloquial Finnish speech into the Lahjoita puhetta (Donate Speech) corpus. The corpus includes over twenty thousand speakers from all the regions of Finland and from all age brackets. The primary goals of the collection were to create a representative, large-scale resource to study spontaneous spoken Finnish and to accelerate the development of language technology and speech-based services. In this paper, we present the collection process and the collected corpus, and showcase its versatility through multiple use cases. The evaluated use cases include: automatic speech recognition of spontaneous speech, detection of age, gender, dialect and topic and metadata analysis. We provide benchmarks for the use cases, as well down loadable, trained baseline systems with open-source code for reproducibility. One further use case is to verify the metadata and transcripts given in this corpus itself, and to suggest artificial metadata and transcripts for the part of the corpus where it is missing.

Anssi Moisio, Mathias Creutz, Mikko Kurimo

Compositional generalisation (CG), in NLP and in machine learning more generally, has been assessed mostly using artificial datasets. It is important to develop benchmarks to assess CG also in real-world natural language tasks in order to understand the abilities and limitations of systems deployed in the wild. To this end, our GenBench Collaborative Benchmarking Task submission utilises the distribution-based compositionality assessment (DBCA) framework to split the Europarl translation corpus into a training and a test set in such a way that the test set requires compositional generalisation capacity. Specifically, the training and test sets have divergent distributions of dependency relations, testing NMT systems' capability of translating dependencies that they have not been trained on. This is a fully-automated procedure to create natural language compositionality benchmarks, making it simple and inexpensive to apply it further to other datasets and languages. The code and data for the experiments is available at https://github.com/aalto-speech/dbca.

Khalid Alnajjar, Mika Hämäläinen, Jörg Tiedemann et al.

Prerecorded laughter accompanying dialog in comedy TV shows encourages the audience to laugh by clearly marking humorous moments in the show. We present an approach for automatically detecting humor in the Friends TV show using multimodal data. Our model is capable of recognizing whether an utterance is humorous or not and assess the intensity of it. We use the prerecorded laughter in the show as annotation as it marks humor and the length of the audience's laughter tells us how funny a given joke is. We evaluate the model on episodes the model has not been exposed to during the training phase. Our results show that the model is capable of correctly detecting whether an utterance is humorous 78% of the time and how long the audience's laughter reaction should last with a mean absolute error of 600 milliseconds.

Anja Virkkunen, Aku Rouhe, Nhan Phan et al.

Public sources like parliament meeting recordings and transcripts provide ever-growing material for the training and evaluation of automatic speech recognition (ASR) systems. In this paper, we publish and analyse the Finnish parliament ASR corpus, the largest publicly available collection of manually transcribed speech data for Finnish with over 3000 hours of speech and 449 speakers for which it provides rich demographic metadata. This corpus builds on earlier initial work, and as a result the corpus has a natural split into two training subsets from two periods of time. Similarly, there are two official, corrected test sets covering different times, setting an ASR task with longitudinal distribution-shift characteristics. An official development set is also provided. We develop a complete Kaldi-based data preparation pipeline, and hidden Markov model (HMM), hybrid deep neural network (HMM-DNN) and attention-based encoder-decoder (AED) ASR recipes. We set benchmarks on the official test sets, as well as multiple other recently used test sets. Both temporal corpus subsets are already large, and we observe that beyond their scale, ASR performance on the official test sets plateaus, whereas other domains benefit from added data. The HMM-DNN and AED approaches are compared in a carefully matched equal data setting, with the HMM-DNN system consistently performing better. Finally, the variation of the ASR accuracy is compared between the speaker categories available in the parliament metadata to detect potential biases based on factors such as gender, age, and education.

Georgios Karakasidis, Tamás Grósz, Mikko Kurimo

It is common knowledge that the quantity and quality of the training data play a significant role in the creation of a good machine learning model. In this paper, we take it one step further and demonstrate that the way the training examples are arranged is also of crucial importance. Curriculum Learning is built on the observation that organized and structured assimilation of knowledge has the ability to enable faster training and better comprehension. When humans learn to speak, they first try to utter basic phones and then gradually move towards more complex structures such as words and sentences. This methodology is known as Curriculum Learning, and we employ it in the context of Automatic Speech Recognition. We hypothesize that end-to-end models can achieve better performance when provided with an organized training set consisting of examples that exhibit an increasing level of difficulty (i.e. a curriculum). To impose structure on the training set and to define the notion of an easy example, we explored multiple scoring functions that either use feedback from an external neural network or incorporate feedback from the model itself. Empirical results show that with different curriculums we can balance the training times and the network's performance.

Anssi Moisio, Mathias Creutz, Mikko Kurimo

Rule-based language processing systems have been overshadowed by neural systems in terms of utility, but it remains unclear whether neural NLP systems, in practice, learn the grammar rules that humans use. This work aims to shed light on the issue by evaluating state-of-the-art LLMs in a task of morphological analysis of complex Finnish noun forms. We generate the forms using an FST tool, and they are unlikely to have occurred in the training sets of the LLMs, therefore requiring morphological generalisation capacity. We find that GPT-4-turbo has some difficulties in the task while GPT-3.5-turbo struggles and smaller models Llama2-70B and Poro-34B fail nearly completely.

Dejan Porjazovski, Anssi Moisio, Mikko Kurimo

Test data is said to be out-of-distribution (OOD) when it unexpectedly differs from the training data, a common challenge in real-world use cases of machine learning. Although OOD generalisation has gained interest in recent years, few works have focused on OOD generalisation in spoken language understanding (SLU) tasks. To facilitate research on this topic, we introduce a modified version of the popular SLU dataset SLURP, featuring data splits for testing OOD generalisation in the SLU task. We call our modified dataset SLURP For OOD generalisation, or SLURPFOOD. Utilising our OOD data splits, we find end-to-end SLU models to have limited capacity for generalisation. Furthermore, by employing model interpretability techniques, we shed light on the factors contributing to the generalisation difficulties of the models. To improve the generalisation, we experiment with two techniques, which improve the results on some, but not all the splits, emphasising the need for new techniques.

Dejan Porjazovski, Tamás Grósz, Mikko Kurimo

Traditional topic identification solutions from audio rely on an automatic speech recognition system (ASR) to produce transcripts used as input to a text-based model. These approaches work well in high-resource scenarios, where there are sufficient data to train both components of the pipeline. However, in low-resource situations, the ASR system, even if available, produces low-quality transcripts, leading to a bad text-based classifier. Moreover, spontaneous speech containing hesitations can further degrade the performance of the ASR model. In this paper, we investigate alternatives to the standard text-only solutions by comparing audio-only and hybrid techniques of jointly utilising text and audio features. The models evaluated on spontaneous Finnish speech demonstrate that purely audio-based solutions are a viable option when ASR components are not available, while the hybrid multi-modal solutions achieve the best results.

Yaroslav Getman, Tamás Grósz, Mikko Kurimo et al.

This paper presents the "Non-native Children's Automatic Speech Assessment" (NOCASA) - a data competition part of the IEEE MLSP 2025 conference. NOCASA challenges participants to develop new systems that can assess single-word pronunciations of young second language (L2) learners as part of a gamified pronunciation training app. To achieve this, several issues must be addressed, most notably the limited nature of available training data and the highly unbalanced distribution among the pronunciation level categories. To expedite the development, we provide a pseudo-anonymized training data (TeflonNorL2), containing 10,334 recordings from 44 speakers attempting to pronounce 205 distinct Norwegian words, human-rated on a 1 to 5 scale (number of stars that should be given in the game). In addition to the data, two already trained systems are released as official baselines: an SVM classifier trained on the ComParE_16 acoustic feature set and a multi-task wav2vec 2.0 model. The latter achieves the best performance on the challenge test set, with an unweighted average recall (UAR) of 36.37%.

Nhan Phan, Anusha Porwal, Yaroslav Getman et al.

We present an efficient end-to-end approach for holistic Automatic Speaking Assessment (ASA) of multi-part second-language tests, developed for the 2025 Speak & Improve Challenge. Our system's main novelty is the ability to process all four spoken responses with a single Whisper-small encoder, combine all information via a lightweight aggregator, and predict the final score. This architecture removes the need for transcription and per-part models, cuts inference time, and makes ASA practical for large-scale Computer-Assisted Language Learning systems. Our system achieved a Root Mean Squared Error (RMSE) of 0.384, outperforming the text-based baseline (0.44) while using at most 168M parameters (about 70% of Whisper-small). Furthermore, we propose a data sampling strategy, allowing the model to train on only 44.8% of the speakers in the corpus and still reach 0.383 RMSE, demonstrating improved performance on imbalanced classes and strong data efficiency.

Mehedi Hasan Bijoy, Dejan Porjazovski, Tamás Grósz et al.

Speech Emotion Recognition (SER) is crucial for improving human-computer interaction. Despite strides in monolingual SER, extending them to build a multilingual system remains challenging. Our goal is to train a single model capable of multilingual SER by distilling knowledge from multiple teacher models. To address this, we introduce a novel language-aware multi-teacher knowledge distillation method to advance SER in English, Finnish, and French. It leverages Wav2Vec2.0 as the foundation of monolingual teacher models and then distills their knowledge into a single multilingual student model. The student model demonstrates state-of-the-art performance, with a weighted recall of 72.9 on the English dataset and an unweighted recall of 63.4 on the Finnish dataset, surpassing fine-tuning and knowledge distillation baselines. Our method excels in improving recall for sad and neutral emotions, although it still faces challenges in recognizing anger and happiness.

Nhan Phan, Mikko Kuronen, Maria Kautonen et al.

Mispronunciation detection (MD) models are the cornerstones of many language learning applications. Unfortunately, most systems are built for English and other major languages, while low-resourced language varieties, such as Finland Swedish (FS), lack such tools. In this paper, we introduce our MD model for FS, trained on 89 hours of first language (L1) speakers' spontaneous speech and tested on 33 minutes of L2 transcribed read-aloud speech. We trained a multilingual wav2vec 2.0 model with entropy regularization, followed by temperature scaling and top-k normalization after the inference to better adapt it for MD. The main novelty of our method lies in its simplicity, requiring minimal L2 data. The process is also language-independent, making it suitable for other low-resource languages. Our proposed algorithm allows us to balance Recall (43.2%) and Precision (29.8%), compared with the baseline model's Recall (77.5%) and Precision (17.6%).

Hemant Kathania, Mittul Singh, Tamás Grósz et al.

This paper describes AaltoASR's speech recognition system for the INTERSPEECH 2020 shared task on Automatic Speech Recognition (ASR) for non-native children's speech. The task is to recognize non-native speech from children of various age groups given a limited amount of speech. Moreover, the speech being spontaneous has false starts transcribed as partial words, which in the test transcriptions leads to unseen partial words. To cope with these two challenges, we investigate a data augmentation-based approach. Firstly, we apply the prosody-based data augmentation to supplement the audio data. Secondly, we simulate false starts by introducing partial-word noise in the language modeling corpora creating new words. Acoustic models trained on prosody-based augmented data outperform the models using the baseline recipe or the SpecAugment-based augmentation. The partial-word noise also helps to improve the baseline language model. Our ASR system, a combination of these schemes, is placed third in the evaluation period and achieves the word error rate of 18.71%. Post-evaluation period, we observe that increasing the amounts of prosody-based augmented data leads to better performance. Furthermore, removing low-confidence-score words from hypotheses can lead to further gains. These two improvements lower the ASR error rate to 17.99%.

Katri Leino, Juho Leinonen, Mittul Singh et al.

Creating open-domain chatbots requires large amounts of conversational data and related benchmark tasks to evaluate them. Standardized evaluation tasks are crucial for creating automatic evaluation metrics for model development; otherwise, comparing the models would require resource-expensive human evaluation. While chatbot challenges have recently managed to provide a plethora of such resources for English, resources in other languages are not yet available. In this work, we provide a starting point for Finnish open-domain chatbot research. We describe our collection efforts to create the Finnish chat conversation corpus FinChat, which is made available publicly. FinChat includes unscripted conversations on seven topics from people of different ages. Using this corpus, we also construct a retrieval-based evaluation task for Finnish chatbot development. We observe that off-the-shelf chatbot models trained on conversational corpora do not perform better than chance at choosing the right answer based on automatic metrics, while humans can do the same task almost perfectly. Similarly, in a human evaluation, responses to questions from the evaluation set generated by the chatbots are predominantly marked as incoherent. Thus, FinChat provides a challenging evaluation set, meant to encourage chatbot development in Finnish.

Tamás Grósz, Mittul Singh, Sudarsana Reddy Kadiri et al.

End-to-end neural network models (E2E) have shown significant performance benefits on different INTERSPEECH ComParE tasks. Prior work has applied either a single instance of an E2E model for a task or the same E2E architecture for different tasks. However, applying a single model is unstable or using the same architecture under-utilizes task-specific information. On ComParE 2020 tasks, we investigate applying an ensemble of E2E models for robust performance and developing task-specific modifications for each task. ComParE 2020 introduces three sub-challenges: the breathing sub-challenge to predict the output of a respiratory belt worn by a patient while speaking, the elderly sub-challenge to estimate the elderly speaker's arousal and valence levels and the mask sub-challenge to classify if the speaker is wearing a mask or not. On each of these tasks, an ensemble outperforms the single E2E model. On the breathing sub-challenge, we study the impact of multi-loss strategies on task performance. On the elderly sub-challenge, predicting the valence and arousal levels prompts us to investigate multi-task training and implement data sampling strategies to handle class imbalance. On the mask sub-challenge, using an E2E system without feature engineering is competitive to feature-engineered baselines and provides substantial gains when combined with feature-engineered baselines.

Mittul Singh, Peter Smit, Sami Virpioja et al.

Character-based Neural Network Language Models (NNLM) have the advantage of smaller vocabulary and thus faster training times in comparison to NNLMs based on multi-character units. However, in low-resource scenarios, both the character and multi-character NNLMs suffer from data sparsity. In such scenarios, cross-lingual transfer has improved multi-character NNLM performance by allowing information transfer from a source to the target language. In the same vein, we propose to use cross-lingual transfer for character NNLMs applied to low-resource Automatic Speech Recognition (ASR). However, applying cross-lingual transfer to character NNLMs is not as straightforward. We observe that relatedness of the source language plays an important role in cross-lingual pretraining of character NNLMs. We evaluate this aspect on ASR tasks for two target languages: Finnish (with English and Estonian as source) and Swedish (with Danish, Norwegian, and English as source). Prior work has observed no difference between using the related or unrelated language for multi-character NNLMs. We, however, show that for character-based NNLMs, only pretraining with a related language improves the ASR performance, and using an unrelated language may deteriorate it. We also observe that the benefits are larger when there is much lesser target data than source data.

Mittul Singh, Sami Virpioja, Peter Smit et al.

In spoken Keyword Search, the query may contain out-of-vocabulary (OOV) words not observed when training the speech recognition system. Using subword language models (LMs) in the first-pass recognition makes it possible to recognize the OOV words, but even the subword n-gram LMs suffer from data sparsity. Recurrent Neural Network (RNN) LMs alleviate the sparsity problems but are not suitable for first-pass recognition as such. One way to solve this is to approximate the RNNLMs by back-off n-gram models. In this paper, we propose to interpolate the conventional n-gram models and the RNNLM approximation for better OOV recognition. Furthermore, we develop a new RNNLM approximation method suitable for subword units: It produces variable-order n-grams to include long-span approximations and considers also n-grams that were not originally observed in the training corpus. To evaluate these models on OOVs, we setup Arabic and Finnish Keyword Search tasks concentrating only on OOV words. On these tasks, interpolating the baseline RNNLM approximation and a conventional LM outperforms the conventional LM in terms of the Maximum Term Weighted Value for single-character subwords. Moreover, replacing the baseline approximation with the proposed method achieves the best performance on both multi- and single-character subwords.

Stig-Arne Grönroos, Sami Virpioja, Mikko Kurimo

There are several approaches for improving neural machine translation for low-resource languages: Monolingual data can be exploited via pretraining or data augmentation; Parallel corpora on related language pairs can be used via parameter sharing or transfer learning in multilingual models; Subword segmentation and regularization techniques can be applied to ensure high coverage of the vocabulary. We review these approaches in the context of an asymmetric-resource one-to-many translation task, in which the pair of target languages are related, with one being a very low-resource and the other a higher-resource language. We test various methods on three artificially restricted translation tasks -- English to Estonian (low-resource) and Finnish (high-resource), English to Slovak and Czech, English to Danish and Swedish -- and one real-world task, Norwegian to North Sámi and Finnish. The experiments show positive effects especially for scheduled multi-task learning, denoising autoencoder, and subword sampling.

Abhilash Jain, Aku Ruohe, Stig-Arne Grönroos et al.

Transformers have recently taken the center stage in language modeling after LSTM's were considered the dominant model architecture for a long time. In this project, we investigate the performance of the Transformer architectures-BERT and Transformer-XL for the language modeling task. We use a sub-word model setting with the Finnish language and compare it to the previous State of the art (SOTA) LSTM model. BERT achieves a pseudo-perplexity score of 14.5, which is the first such measure achieved as far as we know. Transformer-XL improves upon the perplexity score to 73.58 which is 27\% better than the LSTM model.

Stig-Arne Grönroos, Sami Virpioja, Mikko Kurimo

Data-driven segmentation of words into subword units has been used in various natural language processing applications such as automatic speech recognition and statistical machine translation for almost 20 years. Recently it has became more widely adopted, as models based on deep neural networks often benefit from subword units even for morphologically simpler languages. In this paper, we discuss and compare training algorithms for a unigram subword model, based on the Expectation Maximization algorithm and lexicon pruning. Using English, Finnish, North Sami, and Turkish data sets, we show that this approach is able to find better solutions to the optimization problem defined by the Morfessor Baseline model than its original recursive training algorithm. The improved optimization also leads to higher morphological segmentation accuracy when compared to a linguistic gold standard. We publish implementations of the new algorithms in the widely-used Morfessor software package.

Reima Karhila, Anna-Riikka Smolander, Sari Ylinen et al.

For researching effects of gamification in foreign language learning for children in the "Say It Again, Kid!" project we developed a feedback paradigm that can drive gameplay in pronunciation learning games. We describe our scoring system based on the difference between a reference phone sequence and the output of a multilingual CTC phoneme recogniser. We present a white-box scoring model of mapped weighted Levenshtein edit distance between reference and error with error weights for articulatory differences computed from a training set of scored utterances. The system can produce a human-readable list of each detected mispronunciation's contribution to the utterance score. We compare our scoring method to established black box methods.

Umut Sulubacak, Jörg Tiedemann, Aku Rouhe et al.

This paper describes the MeMAD project entry to the IWSLT Speech Translation Shared Task, addressing the translation of English audio into German text. Between the pipeline and end-to-end model tracks, we participated only in the former, with three contrastive systems. We tried also the latter, but were not able to finish our end-to-end model in time. All of our systems start by transcribing the audio into text through an automatic speech recognition (ASR) model trained on the TED-LIUM English Speech Recognition Corpus (TED-LIUM). Afterwards, we feed the transcripts into English-German text-based neural machine translation (NMT) models. Our systems employ three different translation models trained on separate training sets compiled from the English-German part of the TED Speech Translation Corpus (TED-Trans) and the OpenSubtitles2018 section of the OPUS collection. In this paper, we also describe the experiments leading up to our final systems. Our experiments indicate that using OpenSubtitles2018 in training significantly improves translation performance. We also experimented with various pre- and postprocessing routines for the NMT module, but we did not have much success with these. Our best-scoring system attains a BLEU score of 16.45 on the test set for this year's task.

Stig-Arne Grönroos, Benoit Huet, Mikko Kurimo et al.

This paper describes the MeMAD project entry to the WMT Multimodal Machine Translation Shared Task. We propose adapting the Transformer neural machine translation (NMT) architecture to a multi-modal setting. In this paper, we also describe the preliminary experiments with text-only translation systems leading us up to this choice. We have the top scoring system for both English-to-German and English-to-French, according to the automatic metrics for flickr18. Our experiments show that the effect of the visual features in our system is small. Our largest gains come from the quality of the underlying text-only NMT system. We find that appropriate use of additional data is effective.

Stig-Arne Grönroos, Sami Virpioja, Mikko Kurimo

This article describes the Aalto University entry to the WMT18 News Translation Shared Task. We participate in the multilingual subtrack with a system trained under the constrained condition to translate from English to both Finnish and Estonian. The system is based on the Transformer model. We focus on improving the consistency of morphological segmentation for words that are similar orthographically, semantically, and distributionally; such words include etymological cognates, loan words, and proper names. For this, we introduce Cognate Morfessor, a multilingual variant of the Morfessor method. We show that our approach improves the translation quality particularly for Estonian, which has less resources for training the translation model.

Seppo Enarvi, Peter Smit, Sami Virpioja et al.

Today, the vocabulary size for language models in large vocabulary speech recognition is typically several hundreds of thousands of words. While this is already sufficient in some applications, the out-of-vocabulary words are still limiting the usability in others. In agglutinative languages the vocabulary for conversational speech should include millions of word forms to cover the spelling variations due to colloquial pronunciations, in addition to the word compounding and inflections. Very large vocabularies are also needed, for example, when the recognition of rare proper names is important.

Seppo Enarvi, Mikko Kurimo

We present a new tool for training neural network language models (NNLMs), scoring sentences, and generating text. The tool has been written using Python library Theano, which allows researcher to easily extend it and tune any aspect of the training process. Regardless of the flexibility, Theano is able to generate extremely fast native code that can utilize a GPU or multiple CPU cores in order to parallelize the heavy numerical computations. The tool has been evaluated in difficult Finnish and English conversational speech recognition tasks, and significant improvement was obtained over our best back-off n-gram models. The results that we obtained in the Finnish task were compared to those from existing RNNLM and RWTHLM toolkits, and found to be as good or better, while training times were an order of magnitude shorter.