# oto

> oto (otoearth, Inc.) builds full-duplex, two-channel conversational speech datasets in English, Japanese, and Korean, plus turn-taking benchmarks and speech models for voice AI. Full-duplex means each speaker is recorded on a separate channel at the source, so overlap, backchannels, and interruptions are preserved instead of being flattened into one mixed track. oto has released five otoSpeech datasets on Hugging Face, and its hand-annotated corpus is the training set for Sesame's TurnBench turn-taking benchmark.

Key facts:
- Languages: English, Japanese, and Korean, each with two speakers and one speaker on each channel. Japanese covers free conversation, role-play, and scripted dialogue at 48 kHz. Korean is free conversation at 44.1 kHz with ASR transcripts for both speakers. Japanese and Korean are licensed on request
- Five full-duplex English datasets released on Hugging Face (280, 141, 104.94, 20, and 17 hours). The 141-hour set draws partly on the 280-hour set, so the hours do not add up to one total
- Commercial catalog includes otoSpeech 1,125h (English, 3,878 sessions), plus egocentric video, multi-speaker, face-to-face, video, role-play, scripted, Japanese, and Korean conversation
- otoSpeech is the training set for TurnBench, Sesame's open turn-taking benchmark (arXiv:2608.25218); oto researchers co-authored the paper
- otoTurn, oto's end-of-turn detection model, ranks first of seven systems on the otoSpeech evaluation with 47.6% fewer false interruptions than its base model
- otoSpeech has been downloaded by teams at OpenAI, Google DeepMind, Meta, NVIDIA, and other AI labs
- Cited in the literature, including work from Brno University of Technology, Carnegie Mellon, and NVIDIA (arXiv:2605.15442)
- Used in published research and open models, including models from Pipecat and an end-of-turn benchmark from LiveKit; the full list with sources is at https://www.oto.earth/use-cases/
- oto also collects custom speech datasets to specification: languages, task designs, turn-taking labels, multi-party, in-person, and video-grounded setups

## Datasets

- [otoSpeech-full-duplex-280h](https://huggingface.co/datasets/otoearth/otoSpeech-full-duplex-280h): 280 hours, diverse and multi-condition with realistic noise. The pretraining workhorse. CC BY 4.0, access by request on Hugging Face.
- [otoSpeech-full-duplex-processed-141h](https://huggingface.co/datasets/otoearth/otoSpeech-full-duplex-processed-141h): 141 hours, cleaned, US-native speakers. The fine-tuning track. CC BY 4.0, access by request on Hugging Face.
- [otoSpeech-full-duplex-turn-104h](https://huggingface.co/datasets/otoearth/otoSpeech-full-duplex-turn-104h): 104.94 hours, hand-annotated with 17 turn-taking event labels. The TurnBench training set. Gated, research license.
- [otoSpeech-full-duplex-task-oriented-20h](https://huggingface.co/datasets/otoearth/otoSpeech-full-duplex-task-oriented-20h): 20 hours of task-oriented conversation across seven collaborative tasks, with task stimuli, interaction logs, and reference answers. CC BY 4.0, gated.
- [otoSpeech-full-duplex-task-oriented-v2](https://huggingface.co/datasets/otoearth/otoSpeech-full-duplex-task-oriented-v2): 17 hours of task-oriented conversation, adding word games, co-drawing, and AI-assisted tasks. CC BY 4.0.
- [All datasets](https://www.oto.earth/datasets): the full catalog with audio samples.
- [otoSpeech Egocentric](https://www.oto.earth/datasets/egocentric): first-person video and full-duplex conversation of two to four people folding towels, tidying everyday objects, and playing Jenga, across 18 recording conditions.
- [Multi-speaker datasets](https://www.oto.earth/datasets/multi-speaker): English multi-party conversation with three to five people and a separate recording for each speaker.
- [English datasets](https://www.oto.earth/datasets/english): two-speaker casual English conversation, from 141 to 1,125 hours, with commercial and CC BY 4.0 options.
- [Buy otoSpeech 1,125h](https://www.oto.earth/store/otospeech-1125h): self-serve purchase of 100, 250, 500, 750, or 1,125 hours, from $10 to $20 per hour, under a commercial data license. Delivered on Hugging Face.
- [Japanese datasets](https://www.oto.earth/datasets/japanese): two-speaker Japanese free conversation, role-play, and scripted dialogue. Licensing on request. Also available [in Japanese](https://www.oto.earth/ja/datasets/japanese).
- [Korean dataset](https://www.oto.earth/datasets/korean): two-speaker Korean free conversation with per-speaker ASR transcripts. Licensing on request.
- [Video dataset](https://www.oto.earth/datasets/video): two-speaker English conversation with one camera video per speaker and two-channel audio. Licensing on request.

## Product

- [otoTurn](https://www.oto.earth/product/ototurn.html): End-of-turn detection model for spontaneous conversation. First of seven systems on the otoSpeech evaluation.
- otoEnhance: speech enhancement trained on the same two-channel data. Coming soon.

## Research

- [otoSpeech Video](https://www.oto.earth/research/otospeech-video.html): multimodal English conversation with a camera on each speaker and two-channel audio.
- [otoSpeech Face-to-Face](https://www.oto.earth/research/otospeech-face-to-face.html): in-person English conversation with two to five people on 16 microphone channels.
- [otoSpeech 3-Speaker](https://www.oto.earth/research/otospeech-3-speaker.html): English conversation among three people with a separate recording for each.
- [otoSpeech Korean](https://www.oto.earth/research/otospeech-korean.html): full-duplex Korean conversation with ASR transcripts.
- [otoSpeech Japanese](https://www.oto.earth/research/otospeech-japanese.html): Japanese free conversation, role-play, and scripted dialogue, with overlap measured on the samples. Also available [in Japanese](https://www.oto.earth/ja/research/otospeech-japanese.html).
- [otoSpeech Egocentric](https://www.oto.earth/research/otospeech-egocentric.html): the tasks, conditions, and signals of a first-person full-duplex conversation dataset.
- [otoSpeech is the training set for TurnBench](https://www.oto.earth/research/turnbench-training-set.html): the collaboration with Sesame.
- [otoSpeech Task: task-oriented full-duplex dataset](https://www.oto.earth/research/task-oriented-dataset.html): conversation released with the task still attached.
- [otoTurn: end-of-turn detection for spontaneous conversation](https://www.oto.earth/research/ototurn-end-of-turn.html): one change, the training data, cut false interruptions by nearly half.

## Guides

- [How to Buy Full-Duplex Conversational Speech Data](https://www.oto.earth/guides/how-to-buy-full-duplex-speech-data.html): what full-duplex data is, how much published systems used, open and commercial options with licenses, what to check in a sample, and pricing. oto's licensed datasets run from $10 to $100 per hour of audio; the license and labels such as turn-taking annotation add to the price.
- [What Conversation Data Does a Humanoid Robot Need?](https://www.oto.earth/guides/conversation-data-for-humanoid-robots.html): What changes when the listener has a body: far-field and moving microphones, motor noise, several people at once, addressee detection, and talk during physical tasks. What exists today and what to ask for.
- [Egocentric Video and Speech Datasets Compared](https://www.oto.earth/guides/egocentric-video-speech-datasets.html): EgoCom, Ego4D, Ego-Exo4D, EgoVerse, EPIC-KITCHENS, Seamless Interaction, MM-F2F, and otoSpeech Egocentric compared on conversation, per-speaker audio, group size, tasks, labels, and license.
- [Multi-Party Conversation Data for Robots That Join a Group](https://www.oto.earth/guides/multi-party-conversation-data-for-robots.html): Why two-person data does not prepare a robot for a group, how addressee and floor control work with three or more people, and which multi-party benchmarks and datasets exist.
- [Japanese Conversational Speech Datasets for Full-Duplex Models](https://www.oto.earth/guides/japanese-full-duplex-speech-datasets.html): Academic, TTS, web-scale, and commercial Japanese conversation datasets compared on channel layout, role-play, sample rate, license, and how to get them, with the data behind J-Moshi.
- [Korean Two-Speaker Conversational Speech Datasets Compared](https://www.oto.earth/guides/korean-conversational-speech-datasets.html): LDC telephone corpora, KsponSpeech, MagicData, Nexdata, SpeechData.ai, Basis Conversations 1500, and otoSpeech Korean compared on channels, sample rate, and license.
- [How Much Full-Duplex Data Do You Need? What Published Systems Used](https://www.oto.earth/guides/how-much-full-duplex-data.html): How many hours of real two-channel conversation published full-duplex and speech-to-speech systems trained on, next to the rest of their training data.
- [How to Check a Two-Channel Speech Dataset in 30 Minutes](https://www.oto.earth/guides/check-two-channel-speech-dataset.html): A hands-on audit of a two-channel speech sample: identical channels, crosstalk, overlap, true bandwidth, clipping, and transcript alignment, with scripts you can run.
- [Fisher, Switchboard, and CALLHOME for Speech-to-Speech Models: Licenses and Alternatives](https://www.oto.earth/guides/fisher-switchboard-callhome-alternatives.html): What Fisher, Switchboard, and CALLHOME contain, how LDC licenses them, why telephone audio limits speech generation, and the alternatives with their licenses.
- [End-of-Turn Detection: Benchmarks and Training Data](https://www.oto.earth/guides/end-of-turn-detection-benchmarks-data.html): Turn-taking benchmarks and their metrics, open end-of-turn detectors, and the training data behind them, including TurnBench and otoTurn.
- [Turn-Taking Labels: What to Annotate for Voice Agents](https://www.oto.earth/guides/turn-taking-labels.html): The turn-taking events worth labeling for voice agents, how published schemes define them, human versus model labels, and what to ask a vendor.
- [Real vs. Synthetic Dialogue for Speech-to-Speech Training](https://www.oto.earth/guides/real-vs-synthetic-dialogue.html): What published speech-to-speech systems did with synthetic dialogue, what it cannot provide, and how teams mix it with real two-channel conversation.
- [Free Conversation, Role-Play, or Scripted: What Each Teaches a Speech Model](https://www.oto.earth/guides/free-conversation-role-play-scripted.html): What free conversation, role-play, and scripted dialogue each teach a full-duplex or TTS model, with measured overlap from Japanese samples.
- [日本語の会話音声データセット: Full-duplexモデルに使えるもの](https://www.oto.earth/ja/guides/japanese-full-duplex-speech-datasets.html): the Japanese version of the Japanese datasets guide.

## Use cases

- [Use cases](https://www.oto.earth/use-cases/): public research, models, and tools that use otoSpeech, each with a link to its source.

## Contact

- Email: agent@oto.earth
- Hugging Face: https://huggingface.co/otoearth