What is Multilingual Speech?
Speech data covering multiple languages within one project, used for multilingual ASR, cross-lingual transfer, and language identification.
Multilingual speech data is a corpus spanning more than one language inside a single project, usually with a shared design so that the languages can be compared or combined. The value is not the sum of the languages; it is the fact that they were produced against one specification.
The distinction that matters is between a multilingual corpus and several monolingual corpora sold together. If each language was collected separately, with its own guideline and its own quality bar, merging them produces a corpus with an invisible seam, and models trained across it inherit whichever language was produced most loosely.
Uses fall into three groups: multilingual recognition, where one model serves several languages; cross-lingual transfer, where a well-resourced language supplies what a low-resource one lacks; and language identification, where the model has to decide which language it is hearing.
What buyers get wrong
The biggest risk in a multilingual project is uneven quality: annotation teams for different languages work to different standards, and once merged they drag each other down.
The specification detail that decides everything
Speaker count, hours, and recording conditions must be listed per language; a combined total does not cut it.
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 Multilingual Speech looks like
A multilingual corpus is judged language by language. The aggregate is the least informative number in the delivery.
-
Per-language figures, never a combined total
Speaker count, hours and recording conditions listed separately for each language, because an aggregate hides a language with twelve speakers behind one with two hundred.
-
One guideline applied across all languages
With deviations documented rather than left to the local annotation team. A separate guideline per language produces corpora that cannot be aligned to each other.
-
Per-language inter-annotator agreement
Agreement figures are the only evidence that the guideline was applied consistently, and they have to be reported per language rather than pooled across the project.
-
Balance declared rather than assumed
A corpus where one language is eighty percent of the hours behaves differently in training from a balanced one. Neither is wrong as long as the specification says which it is buying.
-
Language labels at the segment level where switching occurs
Utterance-level labels are not enough when a single utterance contains two languages, which is common in real multilingual usage.
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.
-
Mozilla Common Voice
Crowd-sourced read speech in well over a hundred languages, released under a public licence. Breadth is its strength and depth is its weakness: a well-resourced language may have thousands of hours and a low-resource one a few dozen validated hours.
-
Multilingual LibriSpeech
Read audiobook speech in eight European languages, aligned to the same pipeline as LibriSpeech. Useful for cross-lingual comparison precisely because the collection method is held constant across the languages.
-
FLEURS
A parallel speech corpus covering more than a hundred languages, built on the same sentences across all of them. The parallel structure is what makes it valuable for transfer and for translation work, and each language is small.
-
Why breadth is not depth
A language that appears in all three corpora may still have only a few dozen hours of usable audio. Coverage on a list and coverage at training scale are different claims, and the second is what a deployment needs.
Questions to ask a supplier
Multilingual projects fail at the seams between language teams. These questions find the seams.
- Who writes the annotation guideline, and is it one document or one per language?
- How is agreement measured per language, and what were the figures on the last multilingual project?
- Are the language teams in-house, subcontracted, or a mix?
- What happens when a language team proposes a deviation from the shared guideline?
- Can you deliver per-language manifests rather than one combined manifest?
The vague answer in this category is about language balance. A supplier who says the corpus covers forty languages without giving the hours per language has described a language list rather than a corpus. Forty languages at twenty hours each and forty languages at one hour each are different products, and the second one will not train a multilingual model.
A verifiable answer gives hours, speaker counts and domain composition per language, and states which languages share speakers. A speaker who appears in three languages is useful for some work and contaminating for others. Ask for the per-language table in writing, and check one language against the delivered files before accepting the batch.
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.
| Languages covered | Listed individually with ISO codes |
|---|---|
| Speakers per language | Minimum and actual, reported per language |
| Hours per language | Reported separately, never pooled into one figure |
| Guideline structure | One shared document, with deviations recorded as exceptions |
| Agreement figures | Inter-annotator, reported per language |
| Delivery format | Per-language folders with per-language manifests |
The per-language balance column is the attribute that decides what the corpus can do, and it is the one most often summarised away. A single total-hours figure hides the fact that most of the hours sit in two languages. A model trained on such a corpus performs well on those two and poorly on the rest, which is the opposite of what a multilingual deployment needs.
The values come from the deployment language mix. A product launching in six markets needs those six languages at usable depth, not forty at token depth. Write the six down, set a floor per language, and treat any language above the floor as a bonus rather than as a substitute for the ones below it.
A mistake here is expensive because it is discovered late. Multilingual evaluation averages across languages, so a corpus that is strong in a few and thin in many reports a respectable score while failing the markets that were the reason for the project. The remedy is per-language collection, which means going back to the languages that were under-filled.
How Multilingual Speech is collected
The design work happens before any recording, and it determines whether the languages can be used together.
- Guideline written once, then reviewed per language. One annotation guideline is written and reviewed by a native speaker for each language, with deviations recorded as explicit exceptions rather than absorbed silently.
- Per-language recruitment. Contributor pools are assessed separately, because availability varies enormously and the schedule is set by the scarcest language rather than the largest.
- Parallel recording conditions. Where the languages will be compared, the recording setup is held constant across them so that a difference in the data is a difference in the language rather than in the room.
- Per-language annotation and agreement measurement. Each team annotates against the shared guideline, and agreement is measured inside each language before any merging happens.
- Merge with per-language verification. The corpus is assembled only after each language passes its own check, and per-language manifests are delivered alongside.
After recording, each language goes through transcription by native speakers, a script and orthography check, and packaging with per-language metadata covering speaker counts and domain composition. The orthography check is the step that decides whether the languages are actually comparable, since the same language can be written several ways.
The pilot for a multilingual batch is one language recorded end to end rather than a sample from every language, because the failure this category produces is uneven quality between languages and a thin sample from each hides it. Running one language through the full chain shows what the recruitment, the orthography decision and the transcription review actually cost, and those figures set the plan for the rest.
Acceptance is checked per language against the same criteria, with the agreement figure computed separately for each rather than pooled. The specification should state the minimum hours and speaker count per language before recording starts. Where the pilot shows a language cannot reach the floor, the honest outcome is to drop it from the batch rather than to fill it with a smaller number of speakers, because a thin language drags the model for every language it shares parameters with.
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.
-
One language dominates the total
A delivery reported as ten thousand hours may be nine thousand hours of one language and a thousand spread across the rest. The aggregate looks healthy and the per-language picture is what the model will actually experience.
-
Annotation conventions drift between languages
Each language team makes its own decisions on numbers, punctuation and casing. The corpus looks uniform in the manifest and behaves inconsistently in training, and the drift is invisible until cross-lingual evaluation.
-
Recording conditions differ per language
A language recorded in a studio and another recorded over a phone line produce a difference the model reads as language rather than as channel. Where languages will be compared, the conditions have to be held constant.
Why this matters to buyers
Multilingual recognition is the direct driver. A product deployed in a country with several languages, or across a region, cannot run one model per language if real usage mixes them. It needs a model that handles the set, and that requires training material produced under one design.
The second driver is transfer. A language with few speakers and little written material benefits from what a well-resourced language has learned, and cross-lingual transfer is how that happens. It works only when the two were annotated to the same standard, because otherwise the model is transferring from a corpus it cannot align to.
A multilingual model built on an unbalanced corpus is discovered to be unbalanced by the markets it was built for. The users in the thin languages are the ones who notice, and their reports read as a quality problem rather than as a data problem. Diagnosing it takes longer than fixing it would have, because the corpus looked even on paper.
The second cost is the second collection. Languages that were under-filled have to be collected on their own, at a worse unit price than the original batch, because recruitment and project setup are no longer shared. Teams that set a per-language floor before the first batch do not pay this.
A balanced multilingual corpus gives a model that performs evenly across the markets it was built for, which is what a product launch actually needs. The per-language agreement figures mean the team can see where quality is weakest before release rather than hearing about it from the markets afterwards.
The second gain is that the corpus can absorb a new language. Where the orthography decisions and the recruitment channels are recorded per language, adding a market later means repeating a documented process rather than re-deriving it. Programmes that expand by market depend on that record more than on the size of the original batch.
Specification at a glance
| Category | Multilingual Speech |
|---|---|
| Group | Speech data |
| Critical spec | Speaker count, hours, and recording conditions must be listed per language; a combined total does not cut it. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
Multilingual 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.
-
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 Multilingual Speech
Can you quote a single per-hour price for a multilingual project?
No, and a supplier who does is hiding something. Contributor availability, recruitment cost and annotation difficulty differ by language by large factors, and a blended rate means the easy languages are subsidising the hard ones without saying so.
How do you handle a language where the annotation convention is not standardized?
The convention is established during a pilot batch and written into the guideline as a documented deviation from the shared rules. This is common in low-resource languages and it is planned for rather than treated as a defect.
What if one language turns out to be unavailable at the scale we need?
We report that before the project starts rather than after. The per-language contributor assessment happens during quoting, and if a language cannot reach the required speaker count we will say so instead of delivering a thin language inside a healthy total.
Do you provide language identification labels?
Yes, at the segment level. Utterance-level labels are insufficient wherever a single utterance contains more than one language, which is common in real multilingual usage.
Do you have Multilingual 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 Multilingual 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.