What is Speech Translation?
Parallel data pairing source-language audio with target-language translation, used for speech-to-text and speech-to-speech translation models.
Speech translation data is paired audio and translated text: speech in one language with its translation in another, aligned so the correspondence can be learned. It is the training material for systems that translate speech directly rather than transcribing and then translating.
The hard part is alignment. A translation is not a word-for-word mapping, so the correspondence between a passage of source audio and its translated text is at the segment level, and the segmentation decision determines what the corpus can train.
The uses are speech-to-text translation, where the model outputs translated text, and speech-to-speech translation, where it also produces synthesized speech. The second adds prosody to the requirements, because the output has to sound like speech rather than like a transcript read aloud.
What buyers get wrong
Speech translation data needs all three components: audio, source-language transcription, and target-language translation. Miss one and it cannot be trained on. Many suppliers deliver only two.
The specification detail that decides everything
Whether the translation is human or machine translation with human post-editing: cost and quality differ by an order of magnitude.
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 Speech Translation looks like
Speech translation data is judged on its alignment and on the quality of the translation, which are separate axes.
-
Segment alignment documented
Which span of audio corresponds to which span of translated text, at a granularity the specification names.
-
Translation produced by a translator, not by a machine
Machine-translated reference text trains a model toward the errors of the machine that produced it, and the provenance has to be stated rather than assumed.
-
Translation quality assurance described
Whether the translation was reviewed, by whom, and against what criteria, because a translation is a judgement rather than a transcription.
-
Source audio conditions recorded
Read, spontaneous and conversational speech translate differently, and a corpus at one end does not represent the others.
-
Both directions where the pair is used both ways
A corpus built in one direction trains a model that is weaker in the other, which is a coverage decision rather than a defect.
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.
-
MuST-C
English TED talks with translations into a set of European languages, aligned at the sentence level. It is the standard speech translation corpus and it is one domain: prepared public speaking.
-
CoVoST 2
A many-to-many speech translation corpus built on Common Voice, covering dozens of language pairs. Breadth is high, and the read-speech source and the volunteer recording quality set the ceiling.
-
Europarl-ST
European Parliament speeches with translations across a set of languages. Formal register throughout, and the parallel structure makes it valuable for training even though the domain is narrow.
-
Why domain is the limitation
All three are prepared speech by fluent public speakers. Interpreting in a service or medical setting has different timing, disfluency, and vocabulary, and none of the public corpora contain it.
Questions to ask a supplier
Questions that establish how the translation was produced, because that is the part that cannot be checked afterward.
- Who produced the translation, and were they translating from the audio or from a transcript?
- Was the translation reviewed by a second person?
- At what granularity is the audio aligned to the translated text?
- Is the source speech read, spontaneous, or conversational?
- Are both translation directions covered?
The vague answer in this category concerns what was aligned to what. A supplier who says the corpus is aligned without saying whether the alignment was done on audio or on text has described a step without describing its quality. Text-based alignment does not guarantee that the timings correspond to the spoken segments.
A verifiable answer names the language pairs, the direction of each, whether the translation came from a human or a system, and how the segment boundaries were established. Ask for the alignment method in writing, because in this category the alignment is the product and the audio is the raw material.
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.
| Language pair | Named, with direction stated |
|---|---|
| Translation provenance | Human, machine, or post-edited, stated per record |
| Review | Second-person review, with the criteria described |
| Alignment granularity | Utterance, sentence, or segment level |
| Source speech type | Read, spontaneous, or conversational |
| Delivery format | Audio with aligned translated text and per-record provenance |
The direction column decides what the corpus can train, and it is frequently absent. Translation quality differs substantially by direction, and a corpus covering one direction does not establish the reverse. A model trained on a merged corpus performs unevenly, and the unevenness is invisible without the label.
The values come from the deployment rather than from the language pair. Simultaneous interpretation, consecutive interpretation and document translation have different timing constraints and different error tolerances. A corpus built for one cannot serve another, so the use case belongs in the specification.
A mistake here produces a corpus that evaluates well and streams badly. Where the alignment was done on text, the segment timings drift against the audio, and any application that has to produce output while the speaker is still talking inherits that drift. Fixing it means re-aligning on the audio, which is a full pass over the corpus.
How Speech Translation is collected
Translation is a judgement, so the collection design has to fix who makes it and against what standard.
- Source material and language pair selection. Material is chosen for the domain the system will serve, because translation conventions differ by domain.
- Translation by a qualified translator. The translation is produced by a human translator working from the audio or from a verified transcript, and the choice is recorded because it changes the output.
- Review and adjudication. A second translator reviews the work against a stated criterion, and disagreements are adjudicated rather than averaged.
- Alignment. The audio and the translated text are aligned at the granularity the specification names, and the alignment is verified on a sample.
- Quality assurance and delivery. A sample is assessed for translation adequacy and for alignment accuracy, and both results are delivered with the corpus.
After recording, the audio is segmented, the translations are produced and reviewed, the alignment is verified against the audio rather than the transcript, and the files are packaged with per-segment metadata. The alignment verification is the step that separates a usable corpus from a plausible one.
The pilot for speech translation is one language pair recorded and aligned end to end, because the property being tested is whether the alignment holds against the audio rather than against the transcript. A short run shows the segment timings and the review cost per pair, and those figures decide whether the full set of pairs is feasible. Aligning on text alone produces timings that drift, and the pilot is where that shows.
Acceptance is checked on the alignment verified against the audio and on the direction labels for every pair. The specification should state whether the translation is human or system produced and whether each direction is covered before recording starts. Where the pilot fails, the alignment method is changed rather than the tolerance, since a corpus that streams badly cannot be repaired by re-scoring it.
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.
-
Alignment was done on the text, not the audio
Sentence-level alignment produced by matching translated text does not guarantee that the timings correspond to the spoken segments. Where the use needs streaming or simultaneous translation, the misalignment is fatal.
-
One direction was treated as the reverse
Translation quality differs substantially by direction. A corpus that covers English into a language does not establish the reverse, and the manifest often does not say which direction was collected.
-
Interpreter speech was used as if it were the original
Some corpora are built from interpreter output rather than from the speaker. The result contains interpreting conventions that are not present in the source language, and it is usually undocumented.
Why this matters to buyers
Speech translation is where recognition errors and translation errors compound. A system that transcribes and then translates inherits the recognition error before translation even starts, and direct speech translation avoids that step. Training it requires aligned pairs, which are scarce.
The second driver is coverage of spoken registers. Translation models trained on written text translate written text well and spoken language poorly, because spoken language contains disfluency, incomplete sentences and references that written text does not.
A speech translation model built on text-aligned data fails in the live setting it was bought for. The output arrives after the speaker has moved on, and the latency looks like a model problem rather than a data problem. Diagnosing it requires an alignment audit that most teams run only after the first deployment.
The second cost is the review pass. Where the translation came from a system rather than a human, the errors are systematic and the corpus cannot be corrected by sampling. A full human review costs as much as the original collection and is usually the point at which the project gets re-scoped.
A speech translation corpus aligned on the audio gives a model that produces output while the speaker is still talking, which is the only form that serves simultaneous interpretation. The direction labels mean the team knows which pairs are covered rather than discovering a one-way gap in production.
The second gain is that the review cost is known. Human translation is the dominant cost in this category, and a delivery that records how much was system produced and how much was reviewed lets a second batch be planned rather than estimated. That record is what makes the corpus extensible by pair.
Specification at a glance
| Category | Speech Translation |
|---|---|
| Group | Speech data |
| Critical spec | Whether the translation is human or machine translation with human post-editing: cost and quality differ by an order of magnitude. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
Speech Translation 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 Speech Translation
Can you use machine translation for the reference text?
Only where the specification asks for it and the provenance is labelled per record. A machine-translated reference trains the model toward that machine's errors, which is a real effect and has to be a stated choice rather than a shortcut.
Why does alignment granularity matter?
Because a translation is not word-aligned. At utterance level the model learns a coarse correspondence; at segment level it learns finer ones. The cost differs by a multiple, and so does what the corpus supports.
Should the translator work from the audio or from a transcript?
Both are used, and they produce different results. Working from audio preserves spoken properties that a transcript normalizes away; working from a transcript produces more consistent terminology. The choice is recorded so the corpus stays interpretable.
Do you cover both directions?
Per the specification. Translation quality is not symmetric between directions, and a corpus built in one direction trains a model that is noticeably weaker in the other.
Do you have Speech Translation 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 Speech Translation
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.