What is Audiobook?
Continuous long-form narration in audiobook style, steady in intonation and generous in duration, the main source material for high-quality TTS.
Audiobook data is continuous long-form narration in a steady, controlled delivery: the register a professional narrator uses for a book. Its defining property is duration combined with consistency, and consistency is the hard part.
It differs from read speech in scale rather than in kind. A read speech corpus is many short utterances from many speakers; an audiobook corpus is many hours from few speakers, which inverts the speaker distribution and makes it the natural source material for high quality synthesis of a specific voice.
The commercial uses are text-to-speech, where long-form narration supplies prosody across paragraph and chapter boundaries, and long-form narration products themselves, where a model has to sustain a voice across hours without drift.
What buyers get wrong
Consistency is the hardest part of long-form narration: over ten hours, one speaker's pace and timbre drift. Check segment by segment, and budget for re-records instead of assuming one clean pass.
The specification detail that decides everything
Whether timbre must stay consistent across chapters (the same speaker recording on different days). This drives the schedule and the cost.
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 Audiobook looks like
With long-form narration the risk is drift. The corpus is judged on whether the voice at the end is the same voice as at the beginning.
-
Timbre consistency across the whole recording
A narrator recording across multiple days can drift in pace and tone, and the drift is invisible in any single chapter.
-
Chapter boundaries preserved and labelled
Long-form prosody is what distinguishes this from read speech, and the boundaries are where it shows. Losing them turns the corpus back into utterance data.
-
A documented re-record policy
The question is not whether re-records happen but whether they are marked, because a patched passage can differ audibly from its surroundings.
-
Text alignment at chapter and paragraph level
Alignment is what makes the corpus usable for prosody work rather than only for voice cloning.
-
Clean rights on the text as well as the performance
Narration of a published book involves the underlying work, and the chain has to cover it. Original material written for the project removes the problem.
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.
-
LibriVox
Public-domain audiobooks read by volunteers. It is the source material behind LibriSpeech and the largest freely usable audiobook collection, and the volunteer reading quality varies widely.
-
LibriTTS
A re-processed subset of LibriVox at a higher sample rate with cleaner segmentation, built for text-to-speech. Same source, better packaging.
-
Blizzard Challenge datasets
Yearly corpora released for the Blizzard speech synthesis challenge, typically a single speaker reading a book in a studio. Excellent quality and small, and restricted to challenge participants in most years.
-
Why commercial audiobooks are a different matter
Published audiobooks are licensed to their publishers. They are the best material for this category and they cannot be used without a direct agreement, which is why the freely available shelf is volunteer-read and public-domain only.
Questions to ask a supplier
Questions that reveal whether a supplier has managed a long-form narration project rather than a session-based one.
- How is timbre drift monitored across a multi-day recording?
- Are re-recorded passages marked, and how?
- Is the text cleared, or is the narration of original material?
- How is alignment produced at chapter level?
- What is the session schedule, and how does it protect the narrator's voice?
The vague answer in this category concerns the text again, but for a different reason than read speech. Audiobooks are recordings of published works, and a supplier who says the material is licensed without naming the publisher, the territory and the term has described an intention. The performance also carries a separate right belonging to the narrator.
A verifiable answer names the titles, the rights holder for each, the narrator agreement, and the share of the catalogue that is out of copyright. Ask for per-title clearance rather than a blanket statement, because a single uncleared title makes the whole delivery unsellable, and the audit is far cheaper before the batch than after it.
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.
| Total duration | Reported per narrator, not only as a project total |
|---|---|
| Timbre consistency | Monitored, with the method stated |
| Chapter and paragraph labels | Preserved in the delivery |
| Re-record marking | Documented, so patched passages are identifiable |
| Text rights | Cleared, or original material written for the project |
| Delivery format | Long-form audio with chapter-aligned text and a narrator manifest |
The narration style column decides what the corpus can train, and audiobook delivery is more varied than it sounds. Character voices, neutral narration and dramatic reading are different targets, and a corpus that mixes them without labels cannot be used to build a consistent voice.
The values come from the voice you intend to produce and from the text you intend it to read. A model for long-form narration needs sustained prosody over hours, while a model for short prompts needs range. The session length and the text domain columns are where that distinction gets written down.
A mistake here wastes a long recording. Audiobooks are expensive per hour because the narrator is skilled and the sessions are long, so a corpus recorded in the wrong style is the most costly kind of rework in this category. The style has to be settled before the first session, not reviewed after the last.
How Audiobook is collected
Long-form narration is a scheduling problem before it is a recording problem.
- Text preparation and rights clearance. Either original material is written for the project, or the underlying text rights are cleared before narration begins.
- Narrator selection and screening. A narrator is chosen for the register the specification needs, and screened on a sustained sample rather than a short one.
- Scheduled sessions with drift monitoring. Recording is spread across days with limits on session length, and each session is compared against a reference take before the next one starts.
- Chapter-level alignment. The narration is aligned to the text at chapter and paragraph boundaries, and the alignment is checked rather than generated and assumed.
- Consistency verification before delivery. The full recording is reviewed for drift and for audible re-record seams, because both are invisible in a spot check.
After recording, the audio is aligned to the published text, checked against the manuscript for omissions, and packaged with per-title metadata covering the rights holder, the narrator and the edition. The alignment check is what separates a usable corpus from a collection of readings that drift from the text.
The pilot for audiobook narration is a chapter or two recorded by the intended narrator in the intended style, because the two variables that decide the batch are the narrator and the delivery, and neither can be judged from a sample read. A short run shows whether the style matches the specification and whether the narrator can sustain it over a long session. Both are settled before a catalogue is recorded.
Acceptance rests on the alignment between the audio and the published text, checked for omissions on a held-back portion, and on the per-title rights record. The specification should name the territory and the term for each title before recording starts. Where the pilot fails, the narrator or the style is changed rather than the text, since a full catalogue recorded in the wrong style is the most expensive correction in this category.
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.
-
Chapter boundaries were not respected
Segmentation that cuts across chapters breaks the narrative structure that makes audiobook data useful for long-form work. The segmentation scheme belongs in the delivery specification.
-
One narrator presented as many
A multi-book collection read by one voice gives no speaker diversity. Speaker count has to be stated, because it is a different product depending on the answer.
-
Text and audio were not verified against each other
Volunteer narrators occasionally skip or paraphrase. Where the text is delivered alongside, it has to be checked against the recording, and the check is usually skipped because it is tedious.
Why this matters to buyers
Synthesis quality depends on prosody, and prosody at the paragraph and chapter level cannot be learned from isolated sentences. Long-form narration is the material that supplies it, which is why audiobook corpora command a different price from utterance corpora.
The second driver is the product category itself. Long-form listening, whether an audiobook, a podcast or a document read aloud, requires a voice that stays the same for hours. That is a consistency requirement rather than an accuracy one, and consistency is what long-form corpora are built to test.
A narration model built on the wrong style is rejected by listeners before it is rejected by any metric. The complaint is that it sounds like it is reading rather than telling, and that judgement is not fixable in post-processing. The model has to be rebuilt on new audio, and in this category that audio is the most expensive to produce.
The second cost is rights exposure. Audiobook catalogues are owned by publishers who enforce their rights, and a corpus that cannot show a per-title chain becomes a legal problem at the point of sale rather than at the point of collection. Withdrawing a model built on unclear material costs more than the entire collection.
A narration corpus with verified alignment gives a model clean long-form pairs, which is what a voice for extended text needs and what short-prompt corpora cannot provide. The team avoids a forced-alignment pass, and the style labels mean the model can be conditioned rather than averaged.
The second gain is that the catalogue is cleared per title. Rights in audiobooks sit with publishers and vary between titles, and a delivery that records the holder, the territory and the term for each can be extended with new titles without re-auditing the existing ones. That record is what turns a recording project into a catalogue.
Specification at a glance
| Category | Audiobook |
|---|---|
| Group | Speech data |
| Critical spec | Whether timbre must stay consistent across chapters (the same speaker recording on different days). This drives the schedule and the cost. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
Audiobook 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.
-
Podcast Speech
Long-form podcast and interview audio, either solo monologue or two-person conversation, used for long-form speech recognition and speaker modeling.
Questions about Audiobook
Why is consistency harder here than in short recordings?
Because drift accumulates. A narrator recording ten hours across two weeks can shift in pace and timbre by an amount that is imperceptible day to day and obvious when the chapters are heard together. Catching it requires comparing across sessions rather than checking each one.
Do you narrate existing books?
Only where the text rights are cleared. Narration of a published work involves the underlying copyright as well as the performance, and a corpus that does not cover the text cannot be sold for training.
Are re-recorded passages a problem?
They are a fact of long-form production. The problem is unmarked ones, because a patched passage can differ audibly from its surroundings and a model will learn the seam. Marking them makes the corpus honest.
How is alignment produced?
Against the known text, at chapter and paragraph level. Because the text exists before the recording, alignment is a matching problem rather than a transcription one, which is why this category can be aligned more accurately than spontaneous speech.
Do you have Audiobook 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 Audiobook
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.