Tampere University Speech and Language Technology for AI Master's Thesis Guide: 30 ECTS and 2027 Curriculum Transition

The first rule: the 2027 curriculum is not published yet

Tampere University now lists Speech and Language Technology for AI as a 120 ECTS, two-year Master of Science (Technology) specialisation in Computing Sciences and Electrical Engineering. The applicant page combines speech processing, natural language processing, machine learning, linguistics and computer science. It also gives a crucial timing warning: the programme curriculum is under preparation and will be published in the beginning of 2027.

That warning controls how this guide should be used. The current 2026-2027 CSEE curriculum does not yet list this specialisation. Current Tampere courses in speech processing, signal processing, natural language processing and deep learning show the academic environment from which the new programme is being built, but they do not prove the final future credit structure. Do not invent an 80 ECTS major, a compulsory course list or a programme-specific thesis seminar before Tampere publishes them.

What is already stable is the thesis family. Because the degree is Master of Science (Technology), the university-wide Technology thesis process applies: the thesis is 30 ECTS and graded 0-5.

1. What the programme actually covers

The applicant page spans both written and spoken language. For text, Tampere names language-model-based AI systems, machine translation, information retrieval, summarisation, information extraction and question answering. For speech, it describes analysis of what is said, who is speaking, speaking style and emotional state, plus speech compression, transmission, denoising, intelligibility enhancement and synthetic voice generation.

The degree therefore supports very different thesis questions. One student may study automatic speech recognition; another may evaluate a multilingual language model; another may work on speaker recognition, speech enhancement, conversational agents, translation, information retrieval or language-model adaptation. The programme is not simply an LLM degree and not simply an audio-engineering degree.

2. How to use current Tampere course evidence safely

Current COMP.SGN-A02 explicitly includes audio, speech and natural language processing among the fields of signal processing and machine learning. COMP.SGN.120 Introduction to Audio and Speech Processing covers speech-signal properties, speech coding, automatic speech recognition and speech synthesis. Current DATA.ML.200 Deep Learning covers transformers, encoder-decoder structures and supervised and self-supervised learning.

These are useful subject controls because they show real 2026-2027 Tampere teaching in areas named by the new programme. They are not a hidden preview of the future degree structure. When the 2027 curriculum appears, use the new Sisu module and course list as the authority.

3. The thesis remains 30 ECTS

The Technology thesis is an independent academic project, not merely a software build or benchmark exercise. Thirty ECTS does not require one universal method. A Speech and Language Technology for AI thesis may involve corpus analysis, speech-signal processing, machine learning, system implementation, human evaluation, modelling, experimental comparison, literature review or a combination justified by the research question.

A working model is not yet an academic contribution. The thesis needs a defined problem, justified method, credible baseline or comparison, interpretable evidence and a discussion of limitations. A higher benchmark score is useful only when the benchmark, data split, metric and comparison support the claim being made.

4. Turn a broad AI idea into a narrow research question

“Fine-tune an LLM” or “build speech recognition” describes an activity, not a thesis question. Define the language or speaker population, input conditions, model or system boundary, comparison and primary outcome. For translation, specify language pair and domain. For ASR, specify speaker/acoustic conditions and whether the claim concerns unseen speakers or unseen domains. For retrieval or question answering, define the corpus and evidence standard.

Scope is especially important in language AI because data preparation, model training, evaluation, deployment and human testing can each become a large project. Decide which stage contains the research contribution and treat the others as supporting infrastructure.

5. Build the data split before model development

Speech and language studies are vulnerable to leakage. The correct split unit may be a speaker, recording session, conversation, document, author, source website, time period or dataset rather than an individual sentence or audio segment. Randomly splitting windows from the same recording across training and test data can create an unrealistically easy task.

Define training, model-selection and final test data before repeated experimentation. Keep the final test set genuinely final. If data are scarce, grouped cross-validation can help, but the grouping rule should match the generalisation claim.

6. Speech experiments need speaker and recording controls

For speech work, record sampling rate, channel configuration, segmentation, loudness or amplitude processing, microphone or corpus conditions and acoustic environment when relevant. If the thesis claims performance on unseen speakers, speaker identity must not leak across training and final evaluation.

Session leakage matters too. Multiple utterances from the same recording session can share microphone characteristics, room acoustics and speaking condition. A model may appear to recognise linguistic or speaker properties while actually exploiting session-specific cues. Match the split design to the claim.

7. Automatic speech recognition and transcription

For ASR, word error rate and character error rate can be useful, but they should be interpreted carefully. Tokenisation, punctuation, normalisation and language-specific morphology can change the meaning of an error rate. Define the scoring procedure and text normalisation before final evaluation.

Do more than report one aggregate number. Examine deletion, insertion and substitution patterns, difficult speakers, noisy conditions and out-of-domain material when those are relevant. If the system is intended for multilingual use, report language-specific performance rather than hiding weak languages inside one average.

8. Speaker, emotion and paralinguistic analysis

The applicant page explicitly mentions identifying who is speaking, speaking style and emotional state. These tasks can involve sensitive personal or behavioural information. Define the target label, ground truth and collection conditions carefully. Emotion labels are especially difficult because annotators may disagree and context can change interpretation.

For speaker recognition, document whether the task is identification or verification, how enrollment and test data are separated and which metric is appropriate. Avoid claiming biometric robustness from a narrow corpus recorded under one condition.

9. Speech enhancement and synthetic speech

Speech enhancement, denoising and synthesis can be evaluated with automatic quality or intelligibility metrics, but metrics do not always track human perception. If perceptual quality is central, consider whether a listening test or another human-evaluation method is needed and feasible.

For text-to-speech or voice generation, separate intelligibility, naturalness, speaker similarity and prosody. A system can improve one dimension while harming another. Synthetic-voice work also raises consent, impersonation and misuse questions, especially when real speakers are modelled.

10. Natural language processing and text corpora

Text corpora can contain duplicates, templates, leaked labels, personal data or copyrighted material. Document corpus source, version, licence or access conditions, filtering and deduplication. If the test benchmark is widely available online, consider whether the evaluated model may already have encountered it during pretraining.

For classification, retrieval, summarisation or information extraction, choose metrics that reflect the task. Accuracy may hide class imbalance. Exact-match metrics can be too strict for some generation tasks, while overlap metrics can reward fluent but unsupported text. Error analysis should accompany the headline score.

11. Large language models: contamination and reproducibility

LLM research can be difficult to reproduce because provider models, safety layers and APIs change. Record the exact model identifier, date or version where available, prompt template, system instructions, sampling parameters, retrieval configuration and post-processing. If an API changes during the thesis, preserve which experiments used which version.

Benchmark contamination is a major boundary. A high score does not prove generalisation if the model may have seen the benchmark or close variants during pretraining. Where contamination cannot be ruled out, state that limitation and consider newer, private or procedurally generated evaluation data.

12. Fine-tuning and adaptation

If the thesis fine-tunes or adapts a model, document the base model, training data, split, objective, optimiser, learning rate, batch size, number of steps or epochs, parameter-efficient method and random seeds where relevant. Keep the untouched final evaluation separate from training and hyperparameter selection.

A larger model is not automatically a stronger scientific baseline. Compare against credible alternatives and report compute or resource trade-offs when efficiency is part of the claim. If the thesis uses only prompting or retrieval rather than weight updates, describe that honestly rather than calling every adaptation “fine-tuning”.

13. Machine translation, summarisation and generation

Generated text can be fluent while being semantically wrong. Automatic metrics such as BLEU, chrF, ROUGE or embedding-based scores capture only parts of quality. Select metrics appropriate to the task and discuss their limits. For high-stakes factual generation, human or reference-based factuality checks may be more important than surface overlap.

When human evaluation is used, define the rubric, annotator background, blinding, number of items and agreement strategy. Do not present subjective preference scores as objective truth without describing how they were collected.

14. Retrieval and question answering

Retrieval systems should distinguish retrieval quality from answer-generation quality. A weak answer may come from poor retrieval, poor generation or both. Report retrieval metrics separately when the research question requires it, and define the corpus snapshot so another researcher knows what information was available.

For retrieval-augmented generation, log chunking, embedding model, index settings, top-k selection, reranking and prompt construction. If sources change over time, preserve the evidence used for the final experiment.

15. Multilingual, dialect and accent evaluation

Language technology often performs unevenly across languages, dialects, accents and domains. If the thesis claims multilingual or broadly applicable performance, report enough subgroup detail to show where the system succeeds and fails. A macro-average can conceal very weak performance for smaller groups.

Be careful with labels such as “native”, “non-native”, dialect or ethnicity. Use categories that are justified by the data and research question, and avoid turning demographic proxies into unsupported causal explanations.

16. Bias, accessibility and social impact

Speech and language systems affect access to information and communication. Recognition errors may be more costly for some speakers, while generated language can reproduce stereotypes or harmful associations. If fairness or accessibility is relevant, define a measurable property rather than making broad ethical claims without evidence.

A thesis does not need to solve every social problem in AI. It does need to identify foreseeable risks created by its data, evaluation or intended use, and explain which risks were studied and which remain outside scope.

17. Human participants and personal data

Voice is not automatically anonymous, and transcripts can reveal identities, health, location or other sensitive information. If the thesis recruits participants or processes personal data, discuss controller roles, lawful processing, minimisation, storage, access and retention with the supervisor before collection begins.

Public datasets also require care. Downloadability is not proof that every redistribution, model-training use or publication is allowed. Record dataset licences and consent limitations. If company or healthcare data are involved, define the public-thesis boundary before the project depends on restricted material.

18. Reproducibility and experiment tracking

Freeze dataset version, split files, preprocessing, model version, code revision, environment, hyperparameters, prompt templates and evaluation scripts. For stochastic training, report seeds or repeated-run strategy and expected variation. Keep experiment IDs so each table or figure can be traced to an exact run.

For speech pipelines, also version feature extraction, segmentation and normalisation. For LLM APIs, record request parameters and date. Reproducibility may not mean bit-identical outputs, but another researcher should be able to reconstruct what was done and understand why results may vary.

19. Baselines, uncertainty and error analysis

Use baselines that test the thesis claim. A simple n-gram, linear classifier, classical speech method, smaller neural model or existing production system can sometimes be more informative than comparing only against the newest large model. Give baselines comparable tuning effort and the same data split.

Report uncertainty when random seeds, annotators, sampling or finite test sets materially affect results. Then inspect errors. Aggregate metrics tell how much; error analysis helps explain where and why. Keep qualitative examples systematically sampled rather than selecting only dramatic cases.

A useful final control is a claim-to-evidence table. List every major conclusion beside the dataset, experiment, human evaluation, figure or analysis that supports it. Mark engineering observations separately from research claims. If a claim such as “robust”, “multilingual”, “real-time”, “fair” or “human-like” has no explicit evidence row, narrow the language before submission. This simple audit is especially valuable in a 30 ECTS thesis because complex AI pipelines can generate far more outputs than the manuscript can justify academically.

For human evaluation, decide the sampling plan before seeing model outputs. Randomise or blind system identity when appropriate, keep instructions stable, and record how ties, missing ratings and annotator disagreement are handled. If only a small convenience sample is feasible, describe it as such. Do not generalise a preference study beyond the speakers, languages, prompts or domains actually evaluated.

For computational experiments, separate exploratory runs from the final confirmatory protocol. It is reasonable to explore architectures and prompts during development, but the final comparison should use a frozen split, metric, baseline and stopping rule. Preserve failed or neutral results when they explain design decisions; silently discarding them can create an exaggerated impression of certainty.

20. AI use while writing an AI thesis

Tampere’s current AI guidance still applies when the thesis itself studies AI. AI can assist coding, debugging, idea generation or language improvement within current rules, but the student remains responsible for correctness and acknowledgement. AI output is not a primary scholarly source.

Do not upload confidential company text, unpublished speech recordings, personal data, credentials or protected code to external AI services without an approved basis. Verify generated citations, equations, code and technical claims independently.

If the thesis spans both speech and text, make the interface between modalities explicit. Define how transcripts are created, whether ASR errors propagate into downstream NLP, whether timestamps or speaker labels are preserved, and which stage is evaluated independently. Otherwise an apparent language-model improvement may simply reflect a change in transcription quality or preprocessing.

21. ITC.CEE.800 is current, but its future placement is unknown

Current ITC.CEE.800 Tools for theses is 5 ECTS and pass/fail. Its current parts are research methods, research ethics/good scientific practice, information searching and career planning. Tampere recommends it as thesis preparation.

However, the new Speech and Language Technology for AI curriculum is not yet published. This guide therefore does not claim that ITC.CEE.800 is definitely compulsory in the future specialisation. Check the 2027 Sisu structure once Tampere publishes it.

22. Seminar requirements must also be rechecked

Tampere’s Technology thesis process includes seminar/presentation-related activity, but no future SLT-AI-specific seminar code is currently defensible. Do not import an old SPML seminar implementation or assume a separate course merely because the programme is housed in CSEE.

When the new curriculum appears, verify the exact seminar or thesis-course registration, presentation, opposition and attendance requirements in Sisu/Moodle. Programme-specific implementation can change even when the university-wide 30 ECTS thesis remains stable.

23. Supervision plan and research plan

Technology students use a Thesis Supervision Plan. Agree on the research question, scope, supervisors, feedback rhythm, schedule, data access, compute resources and examiner arrangements early. For speech and language work, also record corpus permissions, model access, API budgets, GPU availability and any participant recruitment.

Write a preliminary research plan before large-scale training or data collection. Define the primary claim, baseline, split, metric and analysis strategy first. This reduces the risk of months of compute producing results that cannot answer a clear question.

24. Maturity test, Turnitin and Trepo

The master’s degree includes a maturity test. Under current Technology guidance, the thesis abstract normally serves as the maturity test for international master’s students, while language-history circumstances can change the route.

After supervisor permission, the final thesis goes through Turnitin originality checking. Similarity percentage is not an automatic plagiarism decision; the report must be interpreted. The final thesis is then deposited through Tampere’s Trepo process.

25. Public thesis, PDF/A and confidentiality

Approved master’s theses are public documents and are permanently archived electronically. Prepare a valid PDF/A file and test conversion early, especially when the thesis contains phonetic symbols, mathematical notation, figures, code snippets or multilingual fonts.

Confidential corpora, company prompts, proprietary model weights, restricted transcripts or credentials must remain outside the public manuscript. The public thesis still needs enough method and evidence for academic assessment.

26. Assessment, graduation and the 2027 recheck

The MSc (Technology) thesis uses the 0-5 grading scale; Architecture’s pass/fail exception does not apply. ITC.CEE.800, if it appears in your future study plan, is a separate pass/fail course and does not determine the thesis grade. The normal examiner window is generally 21 days, extended to 28 days where the applicable separate maturity route requires it.

For this particular programme add one extra graduation-planning step: recheck the new curriculum in early 2027. Confirm the final advanced-studies module, compulsory courses, ITC.CEE.800 placement, seminar implementation and any Speech and Language Technology for AI-specific thesis instructions. If the published 2027 curriculum differs from this transition guide, the new curriculum controls.