Switchboard Corpus, Fisher and CALLHOME: Commercial Alternatives

Switchboard, Fisher and CALLHOME are the classic corpora of conversational telephone speech. All three hold two-person phone calls, distributed by the Linguistic Data Consortium (LDC) at the University of Pennsylvania and recorded as two-channel, 8 kHz audio with one speaker per channel. Switchboard and CALLHOME still make up the standard English benchmark for conversational ASR, and nothing else lets you compare results with decades of published work.

They are harder to use in a product. Commercial rights come only with an LDC for-profit membership, most of the recordings date from the 1990s and early 2000s, the audio is narrowband, and coverage is mostly American English. Below: what each corpus contains, how LDC licenses it, and the alternatives, from open corpora to commercial datasets and custom telephone-band recordings.

What Switchboard, Fisher and CALLHOME contain

Switchboard-1 Release 2 (LDC97S62). Texas Instruments collected Switchboard in 1990 and 1991. It holds about 2,400 telephone conversations, roughly 260 hours, from 543 speakers in all areas of the United States. An automated operator placed each call and introduced a topic from a list of about 70, and callers were matched so that no two people talked more than once and nobody discussed the same topic twice. The audio is two-channel 8 kHz mu-law, and the orthographic transcripts were later revised at ISIP.

Fisher English (LDC2004S13 and LDC2005S13). LDC collected Fisher English in 2003. Parts 1 and 2 together hold 11,699 conversations of up to 10 minutes, about 1,960 hours. The platform called participants at the times they gave when registering and paired two people who usually didn't know each other, with a topic from a list that changed every 24 hours. The audio is two-channel 8 kHz mu-law. Moshi, dGSLM and other full-duplex dialogue models later learned turn-taking from these calls.

Fisher's transcripts (LDC2004T19 and LDC2005T19) were made by LDC, BBN and WordWave with speed and economy as the goal, giving up some of the mark-up detail and quality control of earlier, smaller corpora.

CALLHOME (LDC97S42 for American English). CALLHOME is closer to private conversation. Native speakers mostly called family members or close friends, chose their own topics, and talked for up to 30 minutes. The American English set has 120 calls, about 56 hours, but its transcripts (LDC97T14) cover only a contiguous 5- or 10-minute segment of each call.

Companion releases in Egyptian Arabic, German, Japanese, Mandarin Chinese and Spanish hold 100 to 120 calls each, placed from North America to family and friends overseas. All are two-channel 8 kHz audio, published in 1996 and 1997.

Fisher Spanish (LDC2010S01). Released in 2010, it has 819 calls of 10 to 12 minutes, about 163 hours, from 136 native speakers of Caribbean and non-Caribbean Spanish recruited in the continental US and Puerto Rico. It used the same platform-driven setup as Fisher English.

The benchmark. The standard English test set, the 2000 Hub5 evaluation (LDC2002S09), is 40 conversations and about 11 hours: 20 unreleased Switchboard calls and 20 CALLHOME American English calls. That is why papers report separate Switchboard and CALLHOME error rates, often labeled SWB and CH.

How LDC licenses them

Under LDC's standard agreements, what you may do depends on what kind of organization you are and whether you are a member:

• Non-members, not-for-profit members and government members may use LDC data for noncommercial linguistic research and education only. They cannot use it to develop or test products for commercialization.
• A for-profit organization that licenses a corpus as a non-member cannot use it for any commercial purpose. If it joins later, it can gain commercial rights to data it already licensed.
• For-profit members can do commercial technology development with data received while they were members, unless a corpus-specific agreement restricts it. Joining gives commercial rights to data from the membership year, and corpora from earlier, closed years, such as Switchboard and Fisher, can be licensed for commercial use at a reduced fee. Those rights continue after the membership ends.

LDC's Join page listed these membership fees in October 2026:

• For-profit organizations: $34,000 for a standard membership, $40,000 for a subscription.
• Not-for-profit and US government organizations: $2,400 standard, $3,850 subscription.

A standard membership includes 16 corpora published in the membership year, and a subscription includes every corpus published that year. Fees for individual older corpora, Switchboard included, are shown only after login.

The practical consequence: a copy of Switchboard licensed by a university is for research and teaching, and even testing a product on it falls outside those terms.

Why teams look past them

Age. Switchboard was recorded in 1990 and 1991 and Fisher in 2003. Topics and vocabulary reflect their time, and none of the audio comes from the devices your users talk into today.

Narrowband audio. At an 8 kHz sample rate, nothing above 4 kHz survives. That suits a model that will only ever hear phone calls. A model that also hears wideband audio from apps, browsers or meeting tools never sees the upper band in training.

Assigned topics between strangers. Switchboard and Fisher pair strangers on a topic from a list, so the speech is spontaneous but shaped by the prompt and by two people meeting for the first time. CALLHOME is more natural, but small, and only part of each call is transcribed.

Language coverage. The English corpora are American English. Fisher Spanish recruited speakers in the US and Puerto Rico, and CALLHOME's other five languages have 100 to 120 calls each. None of these corpora cover Hindi, Indonesian, Swahili or most of the other languages a global product ships in.

Licensing. Commercial rights require a for-profit membership, a five-figure fee for each membership year, plus a fee for each older corpus you add.

Open conversational corpora, license by license

Several open speech corpora capture natural conversation. Check each license before you train anything you plan to sell. Our list of the best open-source speech datasets covers read-speech sets as well.

AMI Meeting Corpus. 100 hours of English meetings recorded on close-talking and far-field microphones, mostly by non-native speakers. About two-thirds are scenario meetings in which participants play a design team. It is released under CC BY 4.0, which permits commercial use with attribution. It is multi-speaker meeting audio rather than two-person calls, and a standard set for speaker diarization.

CANDOR. More than 1,600 conversations and 850 hours between US adults who were strangers, recorded over video chat in 2020. It is available to researchers on request through BetterUp, and its public pages state no license, so get written terms before any commercial use.

Seamless Interaction. Meta's 2025 dataset has more than 4,000 hours of in-person conversation from over 4,000 participants, with denoised 48 kHz audio in a separate file for each participant. Part of it is improvised with professional actors. The license is CC BY-NC 4.0, which excludes commercial use without permission.

CALLHOME on Hugging Face. TalkBank's diarization version of CALLHOME, covering English, German, Japanese, Mandarin and Spanish, is gated and licensed CC BY-NC-SA 4.0, so it is non-commercial too.

Mandarin. AliMeeting has 118.75 hours of meetings with 2 to 4 participants, recorded on an 8-channel microphone array and headset microphones, under CC BY-SA 4.0. MagicData-RAMC has 180 hours of spontaneous conversation from 663 speakers under CC BY-NC-ND 4.0.

Santa Barbara Corpus of Spoken American English. 60 recordings, about 249,000 words, mostly face-to-face conversation, under CC BY-ND 3.0 US. Whether a trained model counts as a derivative work is a question for your legal team.

The pattern: the open sets that clearly allow commercial use are meetings, and the closest matches to Switchboard's design, paired conversation with one channel per speaker, are non-commercial.

Comparison table

Commercial catalog datasets

Commercial catalogs sell conversational speech with the license settled. SpeechData.ai's catalog has 60 datasets, one per language or dialect, totaling 73,000 hours from 7,250 vetted native speakers. Every dataset is spontaneous two-person conversation recorded in-country on dual-channel audio, one channel per speaker, the layout that makes Switchboard convenient for diarization.

Transcripts are time-aligned and human-verified, and every speaker has age, gender, region and recording device metadata. Every speaker also signed explicit commercial-use consent, and the full consent chain is provided for legal review.

As an American English replacement, the US English conversational dataset has 1,500 hours from 150 speakers at $90 per hour of audio: more than five times the size of Switchboard, though smaller than Fisher. Spanish goes well past Fisher Spanish's 163 hours, with 2,000 hours of Mexican Spanish and 1,000 hours recorded in Spain.

The other CALLHOME languages are covered too: Egyptian Arabic, German, Japanese and Mandarin (recorded in Taiwan), with 500 to 2,000 hours each. Prices run from $60 to $95 per hour of audio and are published on each dataset page.

One caveat: the catalog is general conversation, not telephone audio, so on its own it won't reproduce a narrowband call path. Free samples with matching transcripts and the metadata schema go out on request to a company email. Other vendors also sell conversational and telephony data, and our comparison of speech data providers sets them side by side.

Custom collection, including telephone-band calls

When nothing off the shelf matches, record it. Custom speech data collection covers languages, dialects, domains, devices and acoustic environments outside the catalog, scripted or spontaneous, with a feasibility assessment and a quote within one business day.

If you need telephone audio specifically, look at call center speech data. Native speakers play agent and customer in scenarios written for your industry, improvising rather than reading lines. Each speaker is on a separate channel, captured as narrowband telephony or wideband audio on the devices and channels you scope, such as mobile, landline or VoIP.

Names and account details are fictional, so the recordings hold no real customer data. Transcripts come with speaker turns and timestamps, and projects are quoted individually.

Which one to use

• Research and published comparisons. License Switchboard, Fisher and CALLHOME from LDC and report Hub5 numbers. Nothing replaces them as the benchmark.
• An English telephone model you will sell. Either a for-profit LDC membership plus the corpora you need, or a telephone-band collection recorded to your own call scenarios.
• Wideband conversational ASR, or languages beyond English and Spanish. A commercially licensed dual-channel catalog, topped up with custom collection where coverage runs out.
• Meeting transcription. AMI and AliMeeting are open starting points that allow commercial use.