Your reading path
Start with the right questions.
From principle to practice
Public multilingual references are different source designs
Open the worked exampleTurn a language list into a sourcing specification
Start with the model's deployment setting. A transcription tool for spontaneous Mexican Spanish conversation needs different material from a translation system evaluated on short read utterances. Name the language, locale convention, speaking style, channel and task. Explain whether the project requires general coverage or a specific operational condition.
Separate mandatory coverage from expansion preferences. A small amount of an optional language should not be added to the headline total in a way that obscures a missing required group. Ask for a language-level table containing unique sources, accepted duration, independent speakers, transcript coverage and known limitations.
Code-switching deserves its own specification. A recording-level language tag does not establish which spans switch languages or whether the transcript preserves them. Define the unit of language labeling, the intended script and how borrowed terms, names and ambiguous tokens should be handled. Keep unknown or mixed states visible instead of forcing every segment into a single-language category.
Public multilingual references are different source designs
The official FLEURS card describes a multilingual read-speech research collection. It can help frame cross-language evaluation questions, but read utterances are not interchangeable with spontaneous discussion. Inspect the release and language split that match the experiment.
The MLS release describes audiobook-derived speech across eight languages with release-specific training and evaluation resources. That source style differs from telephone, meetings and podcast interviews. Its scale should not be interpreted as equal speaker or topic coverage in every language.
The CoVoST project concerns speech-to-text translation built around Common Voice material; its translation directions and paired text serve a different task from a monolingual ASR target. NVIDIA's Granary-v2 research provides a newer curation and target-variant reference. These are research comparisons, not Fiund inventory or permission to redistribute the referenced recordings.
- Reference
- FLEURS
- Useful question
- Cross-language read-speech evaluation design
- Do not infer
- Spontaneous dialogue or equal deployment fit
- Reference
- MLS
- Useful question
- Audiobook-derived multilingual ASR preparation
- Do not infer
- Telephone channels or conversational timing
- Reference
- CoVoST
- Useful question
- Speech translation direction and paired targets
- Do not infer
- Every language pair or current maintenance
- Reference
- Granary-v2
- Useful question
- Curation and transcript target variants
- Do not infer
- All processed material is freely available commercial inventory
An explicit capability-status matrix
Download the multilingual requirements CSV. Its three rows are requested examples, not available collections. All source and sample status fields begin as not yet reviewed. The rows distinguish Mexican Spanish, Spain Spanish and English–Mexican Spanish code-switching because these requests create different review questions.
For a real proposal, use a small status vocabulary: requested, source identified, sample reviewed, scope agreed, unavailable or pending. Add a review date and evidence reference to any positive status. A language label copied from a source catalog is not sufficient evidence that the selected recordings meet a particular locale or transcription requirement.
Keep feasibility separate from permission and delivery readiness. A source may contain the desired language but have unresolved use restrictions, incomplete transcripts or unknown speaker duplication. Conversely, a source with fewer hours may be ready for a carefully bounded pilot. The matrix should help the buyer see those differences, not act as a decorative list of languages.
Define transcript and translation conventions per group
Agree how numbers, punctuation, abbreviations, disfluencies and partial words should appear. Some tasks need a readable normalized transcript; others require literal repetitions or events. Where multiple target variants exist, keep their field names and derivation relationships explicit. Do not label a normalized target verbatim merely because it follows the recording's meaning.
For languages with alternative script or spelling conventions, record the choice and reviewer instructions. Transliteration, translation and transcription are different outputs. A translation can be semantically adequate while unsuitable as the reference text for a same-language word-error calculation.
Code-switched spans need bilingual review appropriate to the languages involved. Define what happens when a reviewer cannot resolve a term and how uncertain text is represented. An automatic language detector can support triage, but its output should remain an inferred label until the project's agreed review establishes otherwise. Retain the method and version in the annotation record.
Measure independent coverage and prevent evaluation leakage
Count accepted duration, unique speakers and source sessions separately within every requested group. One recurring presenter can contribute many hours without providing broad speaker coverage. Report the concentration of material by source and speaker where the evidence allows it, and state when identity relationships across publishers are unresolved.
Split related sources before segmentation or target expansion. Alternate edits, translated targets and neighboring excerpts can connect records that appear distinct in a file list. A test set designed for unseen speakers needs a different grouping rule from one designed for unseen topics. State the rule and its limits in the release documentation.
Compare performance by language and recording condition alongside the overall score. A weighted aggregate can conceal a weak low-volume group. Keep sample sizes visible and avoid conclusions from a handful of short clips. The goal of the pilot is to identify whether the source supports the intended analysis, not to manufacture a flattering global average.
A pilot acceptance plan with language-specific reviewers
Select material that represents the required channel and speaking style, including difficult cases. For spontaneous speech, this may include overlap, interruptions, names and incomplete utterances. For read speech, pronunciation, prompt correspondence and recording conditions can be more relevant. The selection method should be recorded so the buyer knows what the pilot can and cannot establish.
Check audio validity, transcript coverage, language labeling and source grouping as distinct layers. Then review task-specific quality with suitable language expertise. A timestamp can be structurally valid while pointing to the wrong spoken phrase; a fluent translation can omit an important qualification. Record errors and disagreement outcomes rather than relying on a single undocumented quality percentage.
Agree how failed rows are corrected, replaced or excluded. Report both proposed and accepted hours for each language. If a required subgroup cannot be supported, update the source-status matrix and proposal instead of retaining the same headline promise with a smaller or different collection underneath it.
Pricing uses language-level scope and accepted units
A multilingual project can combine source access, recording, transcription, translation, metadata review and release preparation. The cost drivers differ across languages and tasks. Compare the supplier's unit with the actual requirement: recorded session-hours, accepted speech-hours, translated words or reviewed segment pairs.
The commercial speech comparison provides dated vendor-specific catalog references. Those rates are not a prediction for a new locale, speaker quota or code-switching collection. A public language catalog and bespoke recruitment are different scopes, even when both are priced using an hour somewhere in the proposal.
For a constructed planning example, two 20-hour locale groups plus 10 hours of code-switching sum to 50requested session-hours. They do not establish50 hours of verified supply, a required minimum purchase or a market budget. Ask which groups can be sampled first, what evidence is available and how rejected material affects the accepted denominator before committing to a larger delivery.
What to send Fiund for a useful review
Send the model task, languages and locales, speaking style, channel conditions, required independent speakers, target duration, transcript conventions and intended use. Include which requirements are mandatory and where a pilot can test alternatives. A brief with unresolved fields is useful when those uncertainties are named.
Fiund can review the source and preparation work needed for a proposal. Availability, sample access, review evidence, pricing and permitted use are confirmed for the selected collection and agreement. This page does not promise every named language, annotation layer or recording condition in current inventory.
For optional speaker attributes, use the speaker demographic metadata guide. For multiple tasks over the same audio, read speech instruction-tuning targets. The general speech sourcing overview helps choose among recognition, generation and interaction requirements before extending a language matrix.
Questions buyers ask
Sources & research notes
Public documentation reviewed 2026-11-13. Provider descriptions and published terms are attributed below. Published facts are attributed to their sources. Worked files are labeled as illustrative or as an inspected editorial source; neither implies commercial dataset clearance.
- FLEURS official dataset cardGoogle · Reviewed 2026-11-13
Primary documentation for the adjacent factual comparison. Release access, source rights and Fiund availability must be reviewed separately.
- Multilingual LibriSpeech official releaseOpenSLR / Meta · Reviewed 2026-11-13
Primary documentation for the adjacent factual comparison. Release access, source rights and Fiund availability must be reviewed separately.
- CoVoST original speech translation projectMeta · Reviewed 2026-11-13
Primary documentation for the adjacent factual comparison. Release access, source rights and Fiund availability must be reviewed separately.
- Granary-v2: multilingual speech curation and target variantsNVIDIA Research · Reviewed 2026-11-13
Primary documentation for the adjacent factual comparison. Release access, source rights and Fiund availability must be reviewed separately.
- Datasheets for DatasetsGebru et al. · Reviewed 2026-11-13
Primary documentation for the adjacent factual comparison. Release access, source rights and Fiund availability must be reviewed separately.
