What is Read Speech?
Recordings of speakers reading specified text, with clear pronunciation and known text, the base material for TTS and ASR.
Read speech is a recording of a speaker reading text that is known in advance. The defining property is that the transcript exists before the audio does, which makes it the cheapest high-quality data to produce and the base material for text-to-speech.
The trade-off is built into the definition. Because the text is known and the speaker is reading it, the speech is unusually clean: even pace, few disfluencies, consistent articulation. That is exactly what a synthesis model needs and exactly what makes a recogniser trained only on read speech fail in the field.
The practical consequence is that read speech is usually one component of a larger plan rather than the whole of it. It supplies the clean signal, and it is paired with spontaneous data when the target is recognition rather than synthesis.
What buyers get wrong
The trap with read speech is that it is too clean. It comes out sounding like a news anchor, and a model trained on it underperforms in real conditions. We collect an extra set at natural speaking rate as a comparison.
The specification detail that decides everything
Studio (low noise floor) or home (real environment). The two differ in both signal-to-noise ratio and price.
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 Read Speech looks like
Read speech is easy to produce and easy to produce badly. The difference is in the text and in the consistency.
-
A text corpus designed for phonetic coverage
What matters is which sounds and sound combinations appear, and how often, rather than the literary quality of the passages. General prose leaves large parts of the sound inventory thin.
-
A stated recording condition with a measured noise floor
Studio or home, with the noise floor documented. Studio without a measured floor is a claim rather than a specification, and the noise floor is what a synthesis model inherits.
-
Pace consistency within a speaker
A speaker who speeds up through a long session produces a corpus that varies by position in the session. Session length is a quality control parameter, not just a scheduling one.
-
Speaker spread matched to the target voice
If the model will speak in a particular age range, region or register, the corpus has to contain it. Recruiting whoever is available produces a voice profile nobody chose.
-
A natural-rate subset where the data will be used for more than synthesis
Reading at a comfortable pace is still slower and more even than speaking. The subset is the cheapest way to add spontaneity without changing the speaker distribution.
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.
-
LibriSpeech
960 hours of read English audiobook speech, derived from LibriVox recordings. It is the most-used read-speech corpus in the field, clean, well-aligned, and entirely one genre and one reading style.
-
LibriTTS
A successor built for text-to-speech, with the same source material re-segmented and re-aligned at a higher sample rate. Better for synthesis, same genre concentration.
-
VCTK
About 110 speakers of English with a range of accents, recorded in a studio. Small, valuable for accent coverage, and not large enough to train a recogniser on its own.
-
Common Voice
Crowd-sourced read speech in well over a hundred languages. Coverage is broad and depth per language varies enormously, from tens of hours to a few minutes.
Questions to ask a supplier
Questions that reveal whether a supplier has thought about the text as well as the audio.
- How was the text corpus constructed, and what phonetic coverage does it achieve?
- What is the measured noise floor, and is it consistent across sessions?
- How is pace monitored during a session?
- When a speaker misreads, is the sentence re-recorded or the take kept?
- Is there a spontaneous subset recorded by the same speakers?
The vague answer in this category concerns the text. A supplier who says the material is read aloud without describing how the text was chosen and cleared has skipped the part that determines whether the corpus can be sold. Read speech carries the rights of whatever was read, and public domain status has to be established per item rather than assumed from the genre.
A verifiable answer names the text sources, the selection rule, the speaker count per text, and the proportion of the text that is out of copyright in the relevant jurisdictions. Ask how phoneme coverage was checked, because a corpus that reads a hundred books may still miss common sound combinations, and coverage is measurable before the recording starts.
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.
| Text source | Purpose-built for phonetic coverage, or general prose |
|---|---|
| Recording condition | Studio or home, with the noise floor stated |
| Sample rate and bit depth | Stated, consistent across the whole corpus |
| Speaker spread | Age, gender and region documented per speaker |
| Natural-rate subset | Included or not, with the share stated |
| Delivery format | Per-utterance audio with the reference text and a manifest |
The reading style column decides what the corpus can train, and it is where read speech is most often mis-sold. Neutral studio narration, expressive audiobook delivery and news reading are different targets. A corpus labelled only as read speech may be any of the three, and a text-to-speech model trained on the wrong one sounds wrong in a way that is hard to diagnose.
The values come from the voice you intend to build. If the product needs a calm assistant voice, the style column says neutral, the pace column says moderate, and the session length column says what a narrator can sustain without fatigue. Those are decisions about the product, and a supplier cannot make them on your behalf.
Getting the style wrong wastes the whole collection. Expressive delivery recorded for a neutral target cannot be flattened afterwards without losing quality, and neutral delivery cannot be made expressive. The audio is what it is, and the only remedy is to record again with a different brief.
How Read Speech is collected
The text is prepared before any speaker is booked.
- Text preparation. The corpus is written or selected for phonetic and prosodic coverage, checked against the target language sound inventory, and split into session-sized blocks.
- Recruitment against the target voice range. Where the model will speak in a particular voice profile, the speakers have to cover it.
- Session recording with pace monitoring. Sessions are kept short enough that pace does not drift, and each block is checked as it is recorded rather than at the end.
- Take selection. Misreads are re-recorded rather than edited, and the selection rule is written down so that the corpus stays consistent across speakers.
- Annotation and verification. Text alignment at the required granularity, plus a check that the delivered audio matches the reference text.
After recording, the audio is aligned to the prompt text, checked for misreadings, and packaged with per-utterance metadata covering the text source, the speaker and the session. The misreading check is the step that decides whether the corpus is a clean read or a collection of small errors that will be learned.
The pilot for read speech is a short run over the text domains the specification names, because the property being tested is how the reader handles the material rather than how the microphone performs. Technical text, proper nouns and numerals are where reading corpora go wrong, and a pilot over a sample of each shows it before the full text set is recorded.
Acceptance rests on the alignment between the prompt text and the audio, checked on a held-back portion, and on pronunciation consistency for the names and terms in the text. The specification should list the terms that must be read consistently before recording starts. Where the pilot fails, the prompt set or the reader brief is revised, since re-recording a full text set 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.
-
Reading style is uniform and unstated
Audiobook narration is one prosody, one pace, and one register. A model trained on it meets conversational speech poorly, and the corpus description usually says "read speech" without saying which kind.
-
Text source is not cleared
Read speech inherits the rights of the text that was read. A corpus built from in-copyright books carries a restriction that the audio alone does not reveal, and it has to be documented.
-
Alignment quality varies inside one delivery
Word-level alignments are often produced automatically. Where the aligner failed, the timings are wrong and nothing in the file says so. Alignment error rate belongs in the delivery report.
Why this matters to buyers
Text-to-speech is the primary consumer, and it is unusually sensitive to recording quality. A synthesis model inherits the noise floor, the room and the microphone of its training data, which is why the studio condition exists as a separate and more expensive product.
The second reason is alignment. Because the text is known in advance, read speech is the easiest material on which to produce accurate time alignment, which makes it the standard base for models that need to know not only what was said but when.
A text-to-speech model built on the wrong reading style is rejected by ear rather than by measurement. Listeners cannot say what is wrong, only that it sounds like an announcement when it should sound like a person, and that verdict is not fixable in post-processing. The model has to be retrained on new audio.
The second cost is rights. Read speech is the category where the underlying text is most easily forgotten, and a corpus that cannot show the source and status of every item becomes unsellable the moment a buyer asks. The remedy is a per-item audit that costs far more after the fact than a selection rule applied before recording.
A read corpus with verified alignment gives a model clean text-audio pairs, which is the input most speech work assumes and few deliveries actually provide. The team avoids a forced-alignment pass on the delivery, and the pronunciation consistency means the model does not learn two readings of the same term.
The second gain is that the text set is reusable. Where the prompts and the domain mix are recorded, a second batch can extend the same text set with new material, and the two batches are directly comparable. Programmes building a voice over several years depend on that continuity more than on the size of any single batch.
Specification at a glance
| Category | Read Speech |
|---|---|
| Group | Speech data |
| Critical spec | Studio (low noise floor) or home (real environment). The two differ in both signal-to-noise ratio and price. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
Read 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.
-
Podcast Speech
Long-form podcast and interview audio, either solo monologue or two-person conversation, used for long-form speech recognition and speaker modeling.
-
Multilingual Speech
Speech data covering multiple languages within one project, used for multilingual ASR, cross-lingual transfer, and language identification.
Questions about Read Speech
Why is read speech cheaper than spontaneous speech?
Because the transcript exists before the recording. Nothing has to be listened to and written down, and the speaker output can be checked against a known reference in real time. The saving is in annotation rather than in recording.
Can we use it for recognition training?
As one component. A recogniser trained only on read speech performs well on read speech and poorly on everything else, because it has never seen a false start or an interruption. It is normally paired with spontaneous data.
What is a natural-rate subset?
A portion of the same speakers recorded speaking rather than reading, at their own pace. It is the cheapest way to add spontaneity to an otherwise clean corpus, and because the voices are the same it does not disturb the speaker distribution.
Do you provide the reference text with the audio?
Yes, as a delivery artifact. The text is the specification for the recording, and a delivery that does not include it cannot be verified against anything.
Do you have Read 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 Read 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.