SCOOT 2.0

Open online course book underpinning Aalto's speech processing teaching: speech production and acoustics, analysis, coding, enhancement, transmission, recognition and synthesis, with interactive Python examples. Licensed CC BY-SA 4.0; designed for self-study and course adoption.

ISCA's official archive of recorded conference content: Interspeech sessions from 2010-2022, Odyssey speaker and language recognition workshops (2010-2016) and invited lectures, streamed free via SuperLectures and YouTube, many with downloadable slides. Coverage currently stops around 2022.

Open-access repository of about 42,000 papers from some 400 ISCA conferences and workshops (1970-2026), including all Interspeech and Eurospeech proceedings plus SSW, Odyssey, SPSC, and Blizzard papers. Searchable by author and topic, with DOIs and BibTeX export.

Free Python DSP textbook with a top-down, code-first approach: decomposing sounds into harmonics, spectra, the FFT, spectrograms, filtering, sampling, and LTI systems. Runnable Jupyter notebooks on GitHub. Free HTML and PDF, Creative Commons licensed; second edition in progress.

Steven Smith's 640-page DSP reference, free online in full: sampling, convolution, the DFT/FFT, and FIR/IIR digital filters, with low-math, intuitive explanations and 500+ figures. Browse chapters or download PDFs. Widely used practical reference.

MIT OpenCourseWare course on signal and system analysis: discrete- and continuous-time LTI systems, Fourier series and transforms, Laplace and z-transforms, sampling, and feedback. Full video lectures, notes, and problem sets with solutions. CC BY-NC-SA. Undergraduate core course.

EPFL four-course specialization on discrete-time signal processing: signals as vectors, the DFT/DTFT, FIR/IIR filter design, sampling, quantization, and applications. Video lectures with Python/Jupyter exercises, taught by Prandoni and Vetterli. Access to course materials requires a paid Coursera enrollment or Coursera Plus (free trial and financial aid available); the instructors' textbook, Signal Processing for Communications, is freely available online at sp4comm.org.

Short YouTube lecture series deriving classic speech features step by step: windowing, the short-time Fourier transform, mel filter banks and MFCCs. Slides are free on kamperh.com, which also hosts a companion dynamic time warping series with slides and a notebook.

YouTube series covering sound fundamentals, time- and frequency-domain features, Fourier and short-time Fourier transforms, mel spectrograms and MFCCs, combining intuition, math and librosa implementations in Python. Code and slides are free on GitHub; the updated v2 code targets current environments, while the legacy branch preserves the code used in the videos.

Python library for audio and music analysis: loading, spectrograms, mel features, MFCCs, pitch tracking, beat detection, and visualization utilities. The foundation of countless speech feature-extraction tutorials. v1.0.0 (first stable-API release) released August 2026; ISC license.

Public course materials on discrete-time signals and systems, correlation, Fourier analysis, sampling, and digital filters, with applications including speech processing. The Fall 2026 course page provides lecture notes, MATLAB demonstrations, and speech waveform examples. Posted materials are free to access; course enrollment and assessment use Purdue systems.

Open textbook with embedded video lectures, exercises, and quizzes. Covers phonetics (articulation, IPA transcription, acoustics) and phonology (features, rules), plus morphology, syntax, semantics, and psycholinguistics. Second edition (2022), CC BY-NC-SA licensed. Introductory undergraduate level, with a Canadian-English focus.

The International Phonetic Association's official clickable chart: click any symbol to hear it (recordings by Esling, House, Ladefoged, Wells) and see its phonetic description, IPA number, and Unicode value. Covers the full consonant, vowel, and diacritic inventory. Free.

Peer-reviewed 'Illustrations of the IPA' in the Journal of the IPA: concise phonetic descriptions of individual languages, each with a segment inventory and a transcribed, recorded passage. A model for describing a language's sounds. The journal is now fully open access.

Audio archive of field recordings from hundreds of languages with word lists, connected speech, and phonetic transcriptions, much of it from Peter Ladefoged's UCLA fieldwork. Primary source material for studying the world's speech-sound inventories. Open to everyone, free download.

Stanford's quarter-length graduate course on spoken language technology, emphasizing dialogue and conversational systems: neural ASR, synthesis, affect detection, digital assistants and multilingual speech. Syllabus, slides and assignments are public; lecture videos are restricted to enrolled students. Individual CS224S lectures are also cited in the original SCOOT Phonetics, Speech Synthesis and Speech Recognition sections.

Five publicly accessible lecture-note PDFs and a bibliography from an ESSLLI 2026 advanced course. Covers common ground, feedback and error handling in spoken dialogue, grounding in language models, multimodal interaction, and incremental and multiparty dialogue. Aimed primarily at PhD students; familiarity with dialogue and pragmatics is helpful. Posted learning materials are free to access.

1,000 hours of 16 kHz read English audiobook speech from LibriVox, segmented and aligned, with standard clean/other train, dev and test splits. CC BY 4.0; direct tar.gz downloads with mirrors. The default first corpus for ASR coursework and baselines.

Crowdsourced read and spontaneous speech under CC0: over 42,000 hours across 295 languages (Scripted Speech v27, September 2026), plus a separate Spontaneous Speech set. Since October 2025 downloads are served exclusively through Mozilla Data Collective (free account). Primary source for low-resource-language ASR training and evaluation.

10,000 hours of transcribed English from audiobooks, podcasts and YouTube, in graded XS-XL training subsets plus dev/test. The GitHub tooling is Apache-2.0, but the audio is a gated Hugging Face dataset: SpeechColab's terms of access (accepted with a free account) restrict audio use to non-commercial research and education. Standard corpus for training and benchmarking large supervised ASR models.

About 50,000 hours of read LibriVox audiobook speech in eight European languages (about 44,500 hours English), CC BY 4.0. The linked Hugging Face dataset streams the seven non-English languages (about 6,000 hours); the English portion and the original archives are downloaded from Meta's servers, listed on OpenSLR (SLR94). Standard train and eval sets for multilingual ASR and increasingly for TTS.

Over 1.1 million hours of multilingual web audio in 100+ languages, released under CC BY 3.0. The v3 release is distinct from YODAS v1 and v2. It includes transcripts and English translations where available, timing information, and audio-quality metadata; downloadable in WebDataset format.

In-the-wild speech for generative models: 101k hours (Emilia, CC BY-NC 4.0) plus 114k hours Emilia-YODAS (CC BY 4.0) in Chinese, English, Japanese, French, German, Korean, with transcripts, WebDataset format. Gated Hugging Face access; check the split licenses carefully.

Sound-quality-restored LibriTTS: 585 hours of 24 kHz English read speech from about 2,400 speakers with normalized text, CC BY 4.0, direct OpenSLR download (SLR141). The standard academic corpus for multi-speaker TTS training before scaling to web data.

Speaker recognition corpora: 1,251 (VoxCeleb1) and 6,112 (VoxCeleb2) speakers, over a million utterances from YouTube interviews, with standard verification trial lists. Audio access via request form on the official Oxford page; no blanket open license, research use.

178 hours of read Mandarin Chinese from 400 speakers across accent regions, with manual transcripts, Apache 2.0, direct OpenSLR download (SLR33). The standard entry corpus and character-error-rate benchmark for Chinese ASR.

n-way parallel read speech in 102 languages (roughly 10-12 hours each), built from FLoRes translation benchmark sentences. CC BY 4.0 on Hugging Face. The standard multilingual ASR/LID/speech-translation evaluation set, used for Whisper, MMS and Open ASR multilingual results.

Benchmarks frozen self-supervised speech representations (wav2vec 2.0, HuBERT, WavLM class) with lightweight task heads across ten tasks: ASR, speaker, emotion, spoken language understanding and more. Evaluated via the open-source s3prl toolkit (now in maintenance mode); public leaderboard and papers.

Multilingual successor to SUPERB: ASR and language identification over 154 languages and 200+ varieties, scored with metrics that emphasize worst-performing languages and accent/dialect robustness. Ran as an Interspeech 2025 challenge; data, baselines and results public.

Collaboratively expanding benchmark of 180 instruction-following tasks across speech, music and general audio, spanning classification, regression and sequence generation, for universal spoken language models (ICLR 2025). Tasks and leaderboard hosted on Hugging Face; evaluation code and task-submission tutorial on GitHub.

Continuously updated leaderboard ranking 80+ open and proprietary ASR systems by word error rate and inverse real-time factor across standard English test sets, plus multilingual and long-form tracks. Evaluation scripts open-sourced on GitHub (huggingface/open_asr_leaderboard); results updated continuously.

Evaluation suite for audio-capable large language models: 50+ datasets covering speech understanding, voice/accent/emotion paralinguistics and audio-scene tasks, with LLM-as-judge scoring (NAACL 2025). Open-source toolkit plus live Hugging Face leaderboard; includes Southeast-Asian language and Singlish extensions.

Benchmark for full-duplex spoken dialogue models: pause handling, backchanneling, turn-taking and interruption management, scored with automatic metrics. v2 (Feb 2026) adds a real-time automated multi-turn examiner; v3 code and data released May 2026, adding tool use under disfluent speech.

Field-defining Wiley textbook (344 pages) on automatically recognizing speaker states and traits — emotion, personality, age, health, intoxication — from speech and language. Covers the taxonomy of paralinguistics, corpora, acoustic features, machine learning and evaluation, with a tutorial on the openEAR toolkit lineage.

Archive of the Computational Paralinguistics Challenge (15 editions, INTERSPEECH 2009–2021 and ACM Multimedia 2022–2023): yearly tasks on emotion, age, personality, sleepiness, intoxication, COVID-19, stuttering and vocalisations, with baseline papers, standard feature sets and winners. Site is plain HTTP; data via organisers.

C++ audio feature extractor, the de facto standard in paralinguistics: computes the ComParE 2016 (6,373 features) and GeMAPS/eGeMAPS acoustic parameter sets used across the challenge literature. pip-installable Python wrapper (opensmile-python). Free for private, research and educational use; commercial use requires a license.

Notebook tutorial for audEERING's public wav2vec 2.0 model predicting arousal, dominance and valence, fine-tuned on MSP-Podcast v1.7; weights on Hugging Face and Zenodo (CC BY-NC-SA 4.0, non-commercial). Companion TPAMI 2023 paper analyses transformer-based SER, cross-corpus generalisation and fairness (arXiv:2203.07378).

Multilingual SER toolkit and benchmark (INTERSPEECH 2024): data preparation plus SpeechBrain training recipes for 32 emotion corpora in 14 languages (294 hours total), with intra- and cross-corpus leaderboards comparing 10 pre-trained speech models. Code and processed metadata are public; some corpora require separate licenses or registration.

Searchable, community-maintained table of 77 speech-emotion corpora (2002–2024) — IEMOCAP, CREMA-D, MSP-Podcast, RAVDESS, RECOLA and more — listing language, size, emotion labels, license and open/restricted access per dataset, with automated link checking. Companion GitHub repository accepts contributions.

Multimodal corpus of roughly 12 hours of dyadic acted and improvised English dialogue from 10 actors, with audio, video, motion capture and transcripts, annotated with categorical emotions and valence/arousal/dominance. Academic access requires a verified institutional email and acceptance of USC's release license; commercial use is not permitted.

Successor to the Odyssey 2024 emotion challenge: categorical (eight classes, Macro-F1) and dimensional (arousal/valence/dominance, CCC) tracks over 324+ hours of naturalistic MSP-Podcast speech. Baseline code on GitHub and the final leaderboards remain online; data requires a free institution-signed academic license.

Curated, peer-reviewed clinical- and professional-topic pages from the field's US professional body, covering aphasia, dysarthria, voice disorders, stuttering and more with assessment, treatment and linked evidence. Free web reference orienting newcomers to clinical terminology and scope of practice.

Guidance resource on captions/subtitles, transcripts, audio description and sign language for media, with the WCAG requirements they satisfy and planning advice. Free, maintained (updated September 2024); defines what quality captioning for Deaf and hard-of-hearing users actually requires.

~18 GB corpus of aligned acoustic and 3D articulatory (EMA) recordings from 8 speakers with dysarthria (cerebral palsy or ALS) plus 7 matched controls, with transcriptions. Free for academic, non-profit use from the Toronto page; also distributed via LDC (LDC2012S02).

Large multi-etiology disordered-speech corpus (Parkinson's, ALS, cerebral palsy, stroke, Down syndrome, stuttering, hearing loss) for training accessible ASR. Access by proposal plus a signed data-use agreement. Backed by Amazon, Apple, Google, Meta and Microsoft; ongoing 2025 releases.

TalkBank repository of audio, transcripts and analysis tools for language in dementia, including the Pitt corpus and the ADReSS (Interspeech 2020) and ADReSSo (2021) Alzheimer's-detection challenge datasets. Free with consortium membership via email registration; students need a faculty sponsor. As of October 2026 the DementiaBank page states it 'is under review for potential modification in compliance with Administration directives', so access terms may change.

Open dataset of ~28,000 three-second podcast clips (plus ~4,000 FluencyBank clips) labeled for five stuttering event types (blocks, prolongations, sound/word repetitions, interjections) by three annotators. Labels and metadata on GitHub (dataset CC BY-NC 4.0); audio must be re-downloaded from the original podcast URLs, and as of October 2026 the Stuttering Is Cool and Strong Voices URLs return 404 (about 6,300 of 28,177 clips), while FluencyBank media require a TalkBank login.

Open-access (CC BY) Frontiers in Language Sciences paper (June 2025) on Google Project Euphonia's multilingual disordered-speech data collection (adding Spanish, French, Japanese, Hindi to English), SLP-based quality review, and a meaning-preservation ASR evaluation metric. Links to released open-source tooling.

Web-based voice-banking pipeline: people with ALS/MND record a sentence inventory (a ~300-sentence Gen3 set or a longer 3,000+ set) to build a personalized synthetic TTS voice for AAC devices. $100 to download a completed voice (free for US ALS patients via Team Gleason); donors and clinicians register free.

Biennial challenge benchmarking speaker anonymization: conceal speaker identity while preserving linguistic content and emotion. The fourth edition (2026) had English and multilingual tracks and concluded with a workshop at the SPSC Symposium co-located with Interspeech 2026 (26 September 2026); baselines, datasets, and evaluation code at github.com/Voice-Privacy-Challenge. Results, system descriptions, and workshop presentation materials are now available on the official site.

Community challenge series for speech spoofing and deepfake detection running since 2015. ASVspoof 5 provides a freely available crowdsourced database of about 2,000 speakers and 32 TTS, voice-conversion, and adversarial attacks (paper: arXiv 2502.08857), with protocols and baselines.

The largest annual conference on speech science and technology, spanning phonetics to neural speech models. Interspeech 2026 was held in Sydney (28 September-1 October 2026); next: São Paulo (29 August-2 September 2027), San Antonio (2028), Athens (2029). Students get reduced registration, volunteer positions, mentoring events, and eligibility for ISCA travel grants.

IEEE's annual flagship signal processing conference; speech and audio are among its largest tracks. ICASSP 2026 (Barcelona, 4-8 May) featured over 4,500 papers, tutorials, and SP Grand Challenges (including URGENT) open to students. Next editions: Toronto 2027, Tokyo 2028. Student rates available.

Two biennial single-track IEEE workshops that alternate yearly: ASRU (2025 edition in Honolulu drew a record 447 submissions) and SLT 2026 (Palermo, 13-16 December). Cover ASR, spoken language understanding, dialog, and speech foundation models. Small scale means posters get real senior-researcher traffic.

ISCA's student-run committee (since 2004). Runs mentoring at Interspeech—topic round-tables (two mentors, 6-8 students) and one-on-one matching (both held in person at Interspeech 2026)—plus the Doctoral Consortium (12th edition, held 26 September 2026 at UNSW before Interspeech 2026), the Speech Pitch podcast, Students Meet Experts panels, and volunteer roles.

Grants for students and early-career researchers attending ISCA-organized or ISCA-supported conferences, workshops, and training schools; cover early registration and possibly accommodation. Require an accepted paper, documentation of other funding sought, and no prior ISCA grant; researchers from low-income countries or special situations may also apply.

ISCA-sponsored summer and winter schools, typically five days, training students and early-career researchers on selected speech topics. ISCA contributes up to 5,000 euros per school budget, mainly for inviting lecturers; attendees can separately apply for ISCA grants through the online portal. The linked page is the call for school organisers; individual schools are announced separately.

Long-running challenge series on speech processing in everyday environments, with data and baselines provided. CHiME-9 (results presented at the HSCMA&CHiME 2026 workshop, Barcelona, May 2026) had two tasks: multimodal conversation recognition from 360-degree video and audio (MCoRec) and conversational hearing enhancement (ECHI). CHiME-10, announced July 2026, has three tasks (ECHI-2, URGENT and SG-TSE) and opens on 4 February 2027.

Third iteration, run as an ICASSP 2026 SP Grand Challenge (September 2025–May 2026; final rankings posted) on universal, robust, generalizable speech enhancement. Track 1: one unified system across diverse distortions, domains, and formats, with a 700-hour curated training set; Track 2: speech quality assessment. Data, baselines and leaderboards remain online; the next URGENT round runs as CHiME-10 Task 2 (opens February 2027).

Annual residential summer program at Johns Hopkins (since 1995): a 2-week summer school followed by a 6-week team research workshop where dozens of international researchers tackle frontier speech and language problems. JSALT 2026 (the 12th Frederick Jelinek Memorial Workshop) ran 15 June-7 August 2026; topics for the next workshop are proposed each autumn. Teams include funded positions for graduate and undergraduate students.

ISCA special interest group joining technical and humanities perspectives on speech security and privacy. Runs the annual SPSC Symposium (6th edition held 26 September 2026 with Interspeech 2026, Sydney, jointly with the VoicePrivacy 2026 workshop), which welcomes challenge-participant papers. SIG participation is free; symposium registration paid.

The IEEE SPS technical committee behind ICASSP speech and language sessions and the ASRU/SLT and Odyssey workshops. Its free newsletter (latest edition: Winter 2025, December 2025) carries committee reports on ICASSP speech-area reviewing, SLTC elections and the ASRU/SLT workshops; the TC pages separately list calls for papers and job postings.