What we heard at Interspeech 2026
24
Interspeech 2026 met in Sydney from 27 September to 1 October under the theme “Speaking Together” [1]. We spent the week at booth S5 and as a full-day sponsor of Indigenous Voices Day. Here is what we brought home: four conversations we kept having, and the research from that same week that puts numbers on them.
1. The next thousand languages are a data problem
The idea we heard most often, in sessions and at the booth, went something like this. Speech technology now works well for around a hundred languages, and the next decade is about the next thousand. The constraint people kept naming was not model architecture. It was data: recordings of real speakers in languages and varieties with little or no public audio, transcribed in the script the language is actually written in.
That changes what “coverage” means. A model can list a language as supported on the strength of a few hours of read speech. A product team launching in that language needs to know how the model performs for the people who will use it, and that takes test audio from those people.
2. Accents are where the averages hide
Accented English came up more than any single language. It is also where the public numbers are starkest. GigaSpeechBench, a real-world benchmark released this summer, tests leading foundation models and commercial APIs on six English accents [2]. On Filipino-accented English, word error rates ran from 10.8% for the best system to 44.7% for the worst. On Scottish English the range was 23.7% to 54.2%. On Chinese-accented English it was 13.0% to 66.1%.
These are not small gaps between similar products. On the same accent, one system gets roughly one word in ten wrong and another nearly one in two. An average across accents hides most of that. The lesson teams kept drawing: choose a model on the accents your users have, measured on audio labelled by accent, not on a headline average.
3. Conversation is harder than speech
The second most common request we heard was for conversational audio: two people talking, interrupting, pausing, overlapping. Voice agents are moving to full duplex, listening and speaking at the same time, and the evaluation work published around the conference is sobering.
A paper posted days before the conference evaluated eleven full-duplex spoken dialogue systems on whether they take or hold the turn when a person would. The best scored 0.47, against a human topline of 0.86 [3]. Another tested audio language models acting as voice agents when a command comes from a bystander rather than the user. The best untuned model held back only 14% of the time [4]. Both benchmarks had to be built from real human interaction, because timing, overlap and who-is-speaking only exist in real conversations.
4. Provenance has moved from the ethics panel to procurement
Provenance-aware corpus design came up in research conversations, not only policy ones. The question has shifted from whether data was collected ethically to whether a team can show it: who was recorded, where, with what consent, and what metadata travels with each clip.
The same week showed why. YODAS v3, released on 24 September, offers more than 1.1 million hours of speech in 147 languages under a CC BY 3.0 licence, gathered by crawling the web [5]. It is a large and useful resource for pretraining. It also shows the trade-off. Scale from crawled audio comes without a consent record per speaker or verified speaker metadata. For a team that has to answer a data-disclosure request line by line, those are different products.
What it means for teams shipping voice
Put the four together and the to-do list is short:
Test on the speakers you will serve: their accents, languages and devices, unscripted.
Evaluate conversation as conversation: two speakers, overlap and real turn-taking, not stitched-together monologues.
Ask for the paper trail, a consent record and verified metadata per speaker, before the data reaches a training run.
Budget for the long tail. The next thousand languages will not arrive as a free download.
Where Silencio fits
This is the work we do. Around 2.5 million contributors in roughly 180 countries record on their own devices, in their own homes, streets and workplaces. Through them we reach about 350 languages, and around 500,000 hours of single-speaker speech are available off the shelf. Each recording carries verified speaker metadata and a consent record for every speaker. Where a language or accent isn’t in the catalogue, we collect it to spec with native speakers in the places it is spoken.
Our public samples are on Hugging Face [6], including accent-labelled English and native-script transcriptions of Yoruba, Hausa and Amharic. If you came by booth S5, thank you. If you didn’t, we’re happy to scope an evaluation set for the languages and accents on your roadmap.
Sources
Interspeech 2026, Sydney, 27 Sep–1 Oct 2026, theme “Speaking Together”. https://linguistlist.org/issues/37/562/ and https://isca-speech.org/event-5462701
Tu et al., “GigaSpeechBench: A Real-World Multilingual Speech-to-Text Benchmark”, arXiv 2606.28884 (v3, Jul 2026), Table 5. https://arxiv.org/abs/2606.28884
Shamsaie et al., “Neither Silence nor Overlap Is Failure: Intent-Conditioned Evaluation of Turn-Taking in Full-Duplex Spoken Dialogue Models”, arXiv 2609.27372, 23 Sep 2026. https://arxiv.org/abs/2609.27372
Zhang et al., “Do Audio LLMs Listen Before They Act? Diagnosing Acoustic-Context Gating in Voice Agents”, arXiv 2609.32536, 26 Sep 2026. https://arxiv.org/abs/2609.32536
Chen et al., “YODAS v3: Over 1 Million Hours of High-Bandwidth, Stereophonic, Multilingual Speech”, arXiv 2609.29448, 24 Sep 2026. https://arxiv.org/abs/2609.29448