Fisher, Switchboard, and CALLHOME for Speech-to-Speech Models: Licenses and Alternatives

What each corpus contains

Each corpus records the two sides of a phone call on separate channels, which is why full-duplex research keeps returning to them. Details below are from the LDC catalog pages.

Taken together, the two Fisher parts hold 11,699 conversations and about 1,960 hours. Fisher's speech and transcripts are separate catalog items (LDC2004T19 and LDC2005T19), as are CALLHOME American English's transcripts (LDC97T14) in the first edition. The CALLHOME series also covers Egyptian Arabic, German, Mandarin Chinese, and Spanish.

The three differ in how natural the talk is. LDC describes Fisher's design as maximizing speaker and vocabulary variety, at the cost of more formal talk between strangers. CALLHOME is intimate and natural, with few speakers and long calls. Switchboard sits between them: strangers, topics, and a robot operator placing the calls.

How LDC licensing works

Every corpus in the table lists the same license on its catalog page: the LDC User Agreement for Non-Members. For more on the institution itself, fullduplex.ai has a profile of LDC.

• Non-members. LDC's obtaining-data page says most of its corpora are available to non-member organizations under research-only licenses, and that the license must be signed and paid for before an order is processed. The non-member agreement limits use "only for non-commercial linguistic education, research and technology development". It also says that if that use results in a commercial product, the user must join LDC as a for-profit member and pay the applicable fees before the product is released. Data may not be redistributed outside the user's research group.
• For-profit members. The for-profit membership agreement licenses the data for research and technology development at the sites the member lists. To the extent the law and any separate user agreements allow, the member may incorporate portions of the data into its own work products, including commercial ones. Redistribution beyond that is still excluded.
• Fees. The posted agreement lists a Standard membership at $34,000 per calendar year, covering up to 16 databases from that membership year, and a Subscription membership at $40,000 per calendar year, covering every database released that year. LDC can amend the fee schedule each year, so confirm the current figure with LDC.
• Membership years. LDC memberships follow the calendar year, and members keep perpetual rights to data acquired during their membership years. Fisher's membership years are 2004 and 2005, and Switchboard's are 1993 and 1997. A company joining today does not get them as part of the current year's allocation, so ask LDC how it licenses older corpora to a new member and at what fee.

Questions to put to LDC in writing before you train a commercial model: whether trained model weights count as a work product that incorporates portions of the data, whether you may release those weights publicly, and which of your sites and contractors the license covers. Separate from the license, check what the speakers agreed to. fullduplex.ai's post on consent and licensing for voice data maps the rules in the US, EU, and Japan.

What telephone bandwidth means for speech generation

A signal sampled at 8 kHz carries nothing above 4 kHz. Fisher, Switchboard, and CALLHOME were captured as 8 kHz mu-law telephone audio, and re-encoding to FLAC (as in the CALLHOME second edition) does not bring the upper band back. For speech recognition this was the target condition. For a model that has to generate speech, it is a gap.

• Generation runs at higher rates. Moshi trains on 24 kHz audio, so Kyutai upsampled Fisher from 8 to 24 kHz with a super-resolution model before training (Défossez et al., 2024). NTPP upsampled Fisher to 16 kHz (Wang et al., 2025). A super-resolution model predicts the missing band. It cannot recover what the speaker's voice actually sounded like above 4 kHz.
• Timing and voice come from different places. Moshi learned conversational timing from Fisher, but its own voice came from synthetic dialogue conditioned on one voice actor. That split works when telephone data only has to teach turn-taking.
• Phone handsets vary. Fisher's transcripts record the types of telephones used. If you train on the audio itself, expect handset and line differences across calls.
• Age and register. Switchboard dates from 1990 to 1991 and Fisher from 2003. Topics were assigned in both. A model trained only on them learns how strangers talked on the phone twenty or more years ago.

If you need to check whether a newer dataset is really wideband, measure its band energy before you buy. A 48 kHz file can still hold 8 kHz audio.

Alternatives and their licenses

These datasets put each speaker on a separate track (or a separate stream per participant). Read each license in full; the summaries below follow what each source states.

Notes on the licenses that matter for a speech-to-speech model:

• CC BY 4.0 allows commercial use with attribution. The otoSpeech sets above use it.
• CANDOR is under the CANDOR Corpus License 2.1. BetterUp's summary allows research, development, and commercial use, but forbids redistribution, sharing derived or synthetic versions, and creating synthetic media of participants. Read the full text before training a model that generates voices.
• DailyTalk is CC BY-SA 4.0, which carries share-alike terms, and its authors describe it as freely available for academic use. It is scripted and acted, so its timing is directed, not spontaneous.
• Basis publishes its terms in a LICENSE file you see after access is granted. Its card says every participant agreed at signup to license their recordings for AI training, and that custom licenses are available.
• Non-commercial sets (Seamless Interaction, MagicHub) suit research and evaluation. Their licenses exclude commercial use.

How to choose

• Research only: Fisher, Switchboard, and CALLHOME under the non-member agreement remain the standard references, and results stay comparable with published work.
• Commercial English model: compare the cost of an LDC for-profit membership plus older-corpus licenses with datasets licensed for commercial use, and weigh the 8 kHz ceiling against wideband audio.
• Other languages: CALLHOME covers six languages, but its Japanese set is about 49 hours of 8 kHz audio. For Japanese, Korean, or other languages at training scale, plan to license or collect two-channel data.
• Any case: pull one conversation and check it for channel separation, bleed, and real bandwidth before you sign.

Where oto fits

oto builds full-duplex conversational speech datasets in English, Japanese, and Korean, with one speaker on each channel.

• Open, CC BY 4.0 data. Four otoSpeech datasets on Hugging Face are under CC BY 4.0, including 280 hours of 48 kHz two-channel English conversation.
• Turn-taking labels. otoSpeech-full-duplex-turn-104h has 17 hand-annotated turn-taking event labels under a research license and is the training set for Sesame's TurnBench.
• Licensed catalog. otoSpeech 1,125h of English, Japanese at 48 kHz, Korean at 44.1 kHz, multi-speaker, and video. Samples are on oto.earth/datasets.
• Pricing. Licensed datasets run from $10 to $100 per hour of audio. The license and labels such as turn-taking annotation add to the price, and custom collection is quoted per project. Contact agent@oto.earth.

Frequently asked questions

Can I use Fisher to train a commercial speech model?

Not under the non-member agreement, which is limited to non-commercial research. A commercial product requires LDC for-profit membership, and you should confirm with LDC how older corpora like Fisher are licensed to members.

How much does an LDC membership cost?

LDC's posted for-profit agreement lists $34,000 per year for Standard (up to 16 databases from the membership year) and $40,000 for Subscription. LDC can change fees each year, so confirm with LDC.

What sample rate is Fisher audio?

Fisher is two-channel mu-law audio at 8 kHz, so it carries no content above 4 kHz. Models that generate wideband speech, such as Moshi, upsampled it before training.

What is the difference between Fisher, Switchboard, and CALLHOME?

Fisher is about 1,960 hours of strangers on assigned topics from 2003. Switchboard is about 260 hours from 1990 to 1991. CALLHOME is calls between family and friends, about 56 hours in American English and 49 in Japanese.

Is there a commercially licensed alternative to Fisher?

Yes. otoSpeech has open two-channel English sets under CC BY 4.0 plus a licensed catalog, and CANDOR's current license allows commercial use with restrictions on redistribution and synthetic media.

Are there two-channel conversation datasets in Japanese?

CALLHOME Japanese has about 49 hours at 8 kHz, and open duplex sets are around 10 hours and non-commercial. oto licenses Japanese two-speaker conversation at 48 kHz.

For more datasets, benchmarks, and models, see the fullduplex.ai catalog.

Related guides

• How Much Full-Duplex Data Do You Need? What Published Systems Used
• How to Buy Full-Duplex Conversational Speech Data
• How to Check a Two-Channel Speech Dataset in 30 Minutes