What is Podcast Speech?
Long-form podcast and interview audio, either solo monologue or two-person conversation, used for long-form speech recognition and speaker modeling.
Podcast speech is long-form spoken content produced for an audience: a monologue, an interview, or a conversation between hosts. The defining property is duration. A single episode can run an hour or more, which makes this the cheapest source of long continuous speech and the hardest to annotate well.
It differs from conversational speech in intent. A podcast is made to be listened to, so the speakers are performing for an audience even when they are talking to each other. That produces a register more fluent and more structured than private conversation and less structured than a script.
Commercial value sits in two places: long-form recognition, where the model has to stay coherent across a long recording, and speaker modelling, where the corpus supplies the same voice across many hours. Both need diarization, and diarization is what makes this category expensive.
What buyers get wrong
Copyright and transcription quality are the two big traps with podcast material. Public podcasts cannot be used commercially as they are; you either license them or record original content.
The specification detail that decides everything
Whether speaker diarization annotation is required. It is the most labor-intensive item in long-form audio.
This is the line item that most often separates a dataset that works from one that gets re-ordered. It belongs in the specification before production starts, not in the delivery review.
How we source it
We do not hold inventory in this category. A requirement comes in, we match it to a producer who can meet the specification, and we agree the quality bar with both sides before recording begins. A pilot batch goes first so misalignments surface early.
Scenarios this covers
Buyers rarely ask for a category by name. They describe a situation, and the situation is what tells us whether this category is the right one. If yours is in the list, name it in the brief and it goes straight into the specification. If it is not, describe it anyway, because it usually means the requirement sits in a neighbouring category.
What good Podcast Speech looks like
The audio is easy to obtain. The annotation is what the dataset is judged on.
-
Diarization at a stated accuracy
With the number of speakers per episode documented. Long-form audio is where diarization errors compound, because a mistake early in an episode propagates through everything after it.
-
Transcript accuracy that survives the crosstalk
Interviews are full of interruption, and a transcript that drops the interrupted passages has removed the parts that distinguish this category from scripted dialogue.
-
Rights covering commercial training
Publicly available is not the same as licensed. A corpus assembled from published episodes cannot be sold without the rights, and the rights are often split between the production company and the hosts.
-
Episode-level metadata
Duration, speaker count, topic, and whether the episode is an interview or a monologue. Without it the corpus cannot be sliced by task or compared across batches.
-
Consistent audio treatment
Loudness normalization and any noise reduction applied uniformly across the corpus, or not at all, with the choice documented so the buyer knows what has been done to the signal.
Public datasets in this category
These are openly documented resources, not products we sell. Each one covers part of the category, and knowing where each stops is usually what decides whether a project needs a collection of its own.
-
Spotify Podcast Dataset
About 100,000 podcast episodes with transcripts, released for research. It is the largest public podcast corpus and it is licensed for non-commercial research use, which rules it out for commercial training.
-
MSP-Podcast
A few thousand episodes annotated for emotion and speaker traits. Valuable for paralinguistic work and small relative to the scale a production model needs.
-
PodcastFillers
A focused corpus for disfluency detection, with fillers and discourse markers annotated. Narrow by design and useful precisely because it annotates what general corpora ignore.
-
Why licensing is the wall
Podcast audio is owned by the producers and often carries music and third-party clips. Public corpora exist for research under restricted terms, and commercial use generally requires going to the rights holders directly.
Questions to ask a supplier
For this category the rights question comes first and the audio question second.
- Is the material original, or licensed from the rights holders, and can the chain be produced in documents?
- What diarization accuracy is achieved, and how is it measured?
- How are interrupted and overlapped passages transcribed?
- What audio processing has been applied, and is it reversible?
- How many speakers per episode, and is that documented per file?
In this category the vague answer concerns the mix. A supplier who says the audio is podcast quality without describing the post-processing has described a listening experience rather than a training corpus. Podcasts are published with music beds, advertisement inserts, compression and level normalisation, and every one of those is a distortion the model will learn.
A verifiable answer states whether the delivered files are the published masters or the pre-production recordings, what processing was applied, and whether the music and advertising segments were removed or labelled. Ask for the proportion of each episode that is speech, because a show with a long musical intro and mid-roll breaks can be a third non-speech by duration.
Technical attributes
The items below are the ones a specification can hold a producer to, and the ones that decide whether a delivery is usable for training rather than only present.
| Episode length | Stated range, long-form by definition |
|---|---|
| Speakers per episode | One, two or more, documented per file |
| Diarization | Required, with a stated accuracy target and a measurement method |
| Transcript granularity | Full verbatim, or lightly edited, stated up front |
| Rights basis | Original production or licensed, with documentation |
| Delivery format | Per-episode audio with time-aligned speaker-labelled transcripts |
The processing state column decides reusability, and it is invisible in a waveform. A master that has been compressed and limited for streaming has lost dynamic range that cannot be restored. A corpus delivered as masters can be used for transcription and for style work, but not for anything that depends on the original dynamics.
The numbers should come from the shows you intend to model. Interview formats, panel discussions and solo monologues differ in speaker count, overlap and turn length. Decide which of those the model will meet, then derive the composition from it rather than accepting a mixed bag described as podcasts.
A mistake here produces a model that transcribes well and sounds wrong. The compression and the music beds teach the model a spectral profile that no other audio shares, and the effect appears as a house sound that cannot be removed afterwards. Recovering the clean audio means going back to the production recordings, which the supplier may no longer hold.
How Podcast Speech is collected
Original production removes the rights problem and adds a production cost. Licensed material reverses that trade.
- Rights resolution first. Either the material is produced for the project, or the licensing chain from each rights holder is documented before any work starts.
- Episode planning. Topic and format are set to produce the speaker and register spread the specification needs, including harder cases such as three-way conversation.
- Recording under consistent conditions. Long-form sessions need a stable setup, because a two-hour episode recorded across changing conditions produces a corpus that varies within itself.
- Diarization and transcription. Speaker-attributed transcription against a guideline defining how crosstalk, laughter and unintelligible passages are handled.
- Accuracy verification on a held-back sample. Diarization is checked against a manually labelled sample rather than assumed, because the error rate is part of what the buyer is paying for.
After acquisition, the episodes are segmented into speech and non-speech, speakers are attributed, and the files are packaged with per-episode metadata covering the show, the format and the processing state. The speech and non-speech segmentation is the step that determines whether the corpus is usable for training rather than only for listening.
The pilot for podcast audio is a set of full episodes rather than clips, because the properties that matter here only exist across an episode: the host and guest structure, the advertising segments and the music stings between sections. A pilot over complete episodes shows whether the segmentation rules hold and whether the sponsor reads can be identified reliably.
Acceptance is checked on the segmentation of speech against music and advertising, and on speaker attribution across the episode. The specification should state how sponsor content is to be treated before collection starts, because whether it is removed, labelled or retained changes both the corpus and the rights position. Where the pilot fails, the rule is rewritten and the pilot run again on the same episodes.
How a delivery goes wrong
The defects below are the ones that recur across this category. Each is decided at collection or annotation time, which is why none of them can be repaired in delivery review.
-
Music beds and ads left in
Podcast episodes contain intros, ads, and music. A corpus that leaves them in teaches a recogniser to transcribe jingles, and stripping them afterwards is a manual process that nobody budgets for.
-
Speaker diarisation is wrong in the multi-host case
Two or three hosts who talk over each other produce diarisation errors that are hard to spot in a transcript and expensive to fix. Agreement on speaker labels should be measured, not assumed.
-
Episode metadata is thin
Genre, recording setup, and host count are usually absent. Without them the corpus cannot be stratified, so a buyer cannot hold out one show to test generalisation.
Why this matters to buyers
Long-form recognition is the direct driver. A model that works on ten-second utterances and degrades after five minutes is unusable for meeting transcription, media indexing, or any application where the recording runs long. Long-form corpora are what expose that and what fix it.
The second driver is speaker modelling. Because a podcast supplies the same voice across many hours under varying conditions, it is unusually good material for learning what is consistent about a speaker and what is not. That property is hard to obtain any other way, because it requires duration.
A model trained on published masters learns the sound of podcast production rather than the sound of speech. It performs well on more podcasts and poorly on everything else, and the gap is easy to miss because the evaluation set was drawn from the same source. The cost is a model that cannot be extended beyond the genre it was trained on.
The second cost is rights and consent. Podcast audio belongs to the producers and the guests, and platform terms do not transfer it. A corpus assembled without a chain to the rights holders is a commercial risk discovered at contract time, and the usual outcome is that the whole collection has to be withdrawn and replaced.
A podcast corpus with reliable segmentation gives a model clean speech without the music and advertising beds that would otherwise be learned as part of the signal. The team avoids a separation pass, and the attribution means the corpus can be filtered by speaker role when a task needs only the host or only the guest.
The second gain is that the rights position is documented per episode. Podcast rights sit with the producer and vary between shows, and a delivery that records the position per item can be extended with new shows without re-checking the ones already cleared. That record is what allows a programme to grow by adding feeds rather than by replacing the corpus.
Specification at a glance
| Category | Podcast Speech |
|---|---|
| Group | Speech data |
| Critical spec | Whether speaker diarization annotation is required. It is the most labor-intensive item in long-form audio. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
Podcast Speech across languages
This category is available in every language we cover. The difficulty changes with the language, which is why each language page documents its own constraints.
Related categories
-
Call Center Speech
Customer-service calls, simulated rather than intercepted, capturing both the agent and the caller side, used for call center QC, intent recognition, and dialogue systems.
-
Conversational Speech
Speech from natural conversation between two or more people, on open or semi-structured topics, used for conversational AI, voice assistants, and small talk models.
-
Singing Voice
Vocal recordings with melody, in both a cappella and accompanied form, used for singing voice synthesis, music information retrieval, and lyric alignment.
-
Speech Commands
Targeted recordings of short command words or phrases, usually with many speakers reading each entry several times over, used for wake words and on-device command recognition.
-
Read Speech
Recordings of speakers reading specified text, with clear pronunciation and known text, the base material for TTS and ASR.
-
Multilingual Speech
Speech data covering multiple languages within one project, used for multilingual ASR, cross-lingual transfer, and language identification.
Questions about Podcast Speech
Can you use published podcasts?
Not without a licence. Public availability is not a licence for commercial training, and the rights are usually held by the production company and sometimes separately by the hosts. We work from original production or from material licensed explicitly for the purpose.
Why is diarization so important here?
Because a long recording with multiple speakers is unusable without it. You cannot slice the corpus by speaker, cannot build a speaker model, and cannot evaluate anything speaker-specific. It is also the most labour-intensive item, which is why it dominates the cost.
How long is a long-form episode?
It depends on the use, but duration is the point of the category, so episodes run well past the length of a normal utterance. The specification should state the minimum, because a corpus of five-minute episodes does not test the same thing as a corpus of hour-long ones.
Do you keep the crosstalk?
Yes, and it is transcribed. Interviews are full of interruption, and a corpus that has cleaned it out has removed the part that distinguishes long-form conversation from scripted dialogue.
Do you have Podcast Speech available now?
No. We do not hold inventory. Everything is produced against a specification. That means the timeline starts when the spec is agreed rather than immediately, and it also means the data matches your requirement instead of approximately matching something already built.
Can this be combined with other categories in one delivery?
Yes, and it is common. A single specification can cover several categories across several languages, with one pilot batch and one quality bar. It usually reduces total cost because contributor recruitment and project setup are shared.
What does the consent documentation cover?
Signed contributor consent stating the intended use, the collection methodology, and the annotation guideline the data was produced under. For projects involving EU data subjects the chain is built to satisfy GDPR, including the transfer mechanism. The terms for moving the data across a border.
Request Podcast Speech
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.