otoSpeech 1,125h

What is in each pack

Smaller packs are a random selection of sessions, so every pack has a similar mix of voices. Larger packs add speakers and topics.

Prices are in US dollars before tax. Shares are of the speakers in each pack. A few speakers did not say. Every pack is contained in each larger one.

Specifications

What it is for

• Full-duplex and speech-to-speech modelsTurn-taking, overlap, and backchannels, with each voice clean on its own channel.
• Conversational speech generationNatural timing and prosody from unscripted talk between two people.
• Speech recognition and understandingSpontaneous speech with diverse accents from around the world, centered on the US.

Compared with other English conversation datasets

otoSpeech 1,125h is full-duplex, recorded at 48 kHz, and the lowest price per hour of any listed dataset. It is also the only one you can buy and download today, with no quote and no sales call.

From each provider's public page or listing, checked November 2026. Full-duplex means each speaker is on their own channel; single means both voices share one track.
* Estimate: commercial use of Fisher needs an LDC for-profit membership, about $42,000 in total for about 1,960 hours.

otoSpeech 1,125h and Open Yap 1K

The closest dataset in format: both are 48 kHz with one speaker per channel.

What each session includes

One stereo FLAC at 48 kHz and 16-bit, with one speaker on each channel, and one JSON file. Sessions are packaged as WebDataset shards.

Session

ID, duration, language, and prompt type (an open topic or a specific question or task)

Topic

The topic and the hints shown to both speakers

Speakers

Channel, age, gender, nationality, birth and residence country, first and second language, dialect, accent, education, and occupation. Each field is self-reported and may be empty.

Score

An engagement score for the session

Redactions

Time ranges muted for privacy

Transcripts and turn-taking labels can be added on request.

Example

The JSON for one session. The fields are real; the values are made up for this example.

{
  "session_id": "8c2f41d07b9e4a6f9e3b1d5a7c0e2f94",
  "duration": 931.4,
  "language": "en",
  "session_type": "random-chat",
  "session_subtype": "",
  "session_instruction_metadata": {
    "topics": [
      {
        "topic": "Talk freely about AI.",
        "hints": [
          "Do you use AI tools at work or school?",
          "What would you never let AI decide for you?",
          "What changed for you in the last year?"
        ],
        "product_name": "",
        "product_url": ""
      }
    ]
  },
  "redacted_segments": [
    { "start_sec": 212.4, "end_sec": 213.1 }
  ],
  "speakers": [
    {
      "speaker_id": "4b1e9a72c3d84f05a6e1b2c3d4e5f607",
      "channel": 0,
      "age": 29,
      "name": "",
      "gender": "female",
      "nationality": "American",
      "birth_country": "United States",
      "dialect": "Midwest",
      "first_language": "English",
      "second_language": "Spanish",
      "education_level": "Bachelor's",
      "mbti": "",
      "job": "Teacher",
      "residence_country": "United States",
      "accent_group": "American",
      "accent": "American"
    },
    {
      "speaker_id": "e07d6c5b4a39281f0e9d8c7b6a594837",
      "channel": 1,
      "age": 34,
      "name": "",
      "gender": "male",
      "nationality": "Nigerian",
      "birth_country": "Nigeria",
      "dialect": "",
      "first_language": "English",
      "second_language": "Yoruba",
      "education_level": "Master's",
      "mbti": "",
      "job": "Software engineer",
      "residence_country": "United States",
      "accent_group": "African",
      "accent": "Nigerian"
    }
  ],
  "scores": { "engagement": 4.8 }
}

Load the data

Python with the standard library and soundfile. Each shard holds about six sessions.

import io
import json
import tarfile

import soundfile as sf

# Each session is a FLAC (one speaker per channel) and a JSON with the same name.
def sessions(path):
    pending = {}
    with tarfile.open(path) as tar:
        for member in tar:
            if not member.isfile():
                continue
            key, ext = member.name.rsplit(".", 1)
            pending.setdefault(key, {})[ext] = tar.extractfile(member).read()
            if {"flac", "json"} <= pending[key].keys():
                files = pending.pop(key)
                audio, sr = sf.read(io.BytesIO(files["flac"]))
                yield json.loads(files["json"]), audio, sr

for meta, audio, sr in sessions("shard-000000.tar"):
    a, b = audio[:, 0], audio[:, 1]  # channel 0 and channel 1
    print(meta["session_id"], meta["session_type"], sr, round(meta["duration"]), "s")
    print([(s["channel"], s["accent"]) for s in meta["speakers"]])
    break

Topics and speakers

From the metadata of all 3,878 sessions and 565 speakers.

Audio

What a full-duplex model needs to hear, measured across the set.

Consent and privacy

• Every session was recorded with both speakers' consent, including consent to commercial use and to training AI models.
• Speakers are identified by random IDs. Names are not included.
• If a speaker withdraws consent, we email the licensee the IDs of the affected sessions, and the licensee deletes them within 30 days.
• Age, gender, accent, and the other profile fields are self-reported, and each one is optional.
• Some spans were muted for privacy. Each muted span is listed in the session's metadata.

Access

1. Choose hoursPick a pack, then enter the licensee and your Hugging Face username.
2. PayPay by card, or by bank transfer against an invoice. The license is issued to the licensee you entered.
3. DownloadYour Hugging Face account gets access to the pack right after payment.
4. Join the groupWe email your invite to oto's private buyers' group when it opens.

Full terms: Data License, Terms of Sale, Privacy Notice, 特定商取引法に基づく表記.

Read next

• How to buy full-duplex conversational speech data
• How to check a two-channel speech dataset in 30 minutes
• How much full-duplex data do you need?
• All English conversation datasets

Citation

@misc{otospeech1125h,
  title        = {otoSpeech 1,125h: Full-Duplex English Conversational Speech},
  author       = {otoearth},
  year         = {2026},
  howpublished = {\url{https://www.oto.earth/datasets/english}},
  note         = {Commercial license}
}