You specify the languages, speakers, and format. We scope recruitment, recording, transcription, QA, rights, and delivery in a written project plan.
The rest is proof: who spoke, how it was recorded, the rights attached, and a transcript your model can trust.
Spirelight delivers both.
“I’d like to change my booking to Friday morning.”
We source contributors by language, dialect, age, gender, region, device, or other project criteria. Speakers record through browser-based tools with prompts, consent, audio checks, and project guidelines built into the workflow.
Review progress during production and receive validated batches with audio, transcripts, metadata, and manifests. Your team can test early batches and adjust the collection before the full dataset is finished.
Need Danish conversations, German call-center speech, French dialect coverage, wake words, commands, or audio plus video recordings? A written brief can define the workflow from speaker recruitment to final delivery: speech data collection services for custom recording requirements, and audio annotation services for structuring agreed labels and transcripts.
Synthetic interface example only. Names, locations, ages, recording status, and audio specifications are fictional and are not customer data, live availability, or operating metrics.
A collection can include monologues, dialogues, wake words, commands, scripted prompts, roleplays, and natural conversations, with speaker criteria and feasibility confirmed in the written scope.
An annotation scope can include machine-assisted or human-reviewed transcripts, timestamps, speaker labels, domain terminology, and rules matched to the agreed model requirements.
The delivery plan can package audio, transcripts, metadata, consent references, QA notes, and manifests in the formats agreed with your engineering team.
Four reasons teams use Spirelight for custom speech data collection.
Define the speakers you need by language, dialect, region, age, gender, device, environment, or other project criteria.
Contributor recruitment across 50+ languages and 30+ markets.
We can run remote, on-site, or studio-style sessions using defined microphones, devices, rooms, scripts, and acoustic requirements.
From single-speaker sessions to multi-day collection projects.
Your team can test early deliveries, identify gaps, and update speaker targets, prompts, or guidelines before the full dataset is complete.
Mid-project adjustments without restarting production.
We regularly recruit and review speakers in markets where off-the-shelf datasets are limited, including Nordic languages, regional dialects, and smaller European language varieties.
Recruiters and reviewers across Denmark, Sweden, Norway, Finland, and Iceland.
Your speech data partner for
fine-tuning and language expansion.
A written project can coordinate contributor recruitment, recording workflows, transcription, QA, and delivery for a specified language-expansion brief.
The workflow is similar across projects, but the speakers, prompts, recording conditions, annotations, and deliverables change with each use case.
Collect commands, activation phrases, device instructions, and short utterances across languages, accents, microphones, and environments.
Build language, accent, and dialect coverage for speech recognition, synthetic voice, and speech evaluation datasets.
Collect conversations, roleplays, customer service scenarios, emotional speech, and domain-specific interactions for more natural voice systems.
Spirelight combines commercial project design, recruitment operations, platform engineering, transcription workflows, QA, and delivery management in one team.
Andreas works with clients to turn model requirements into concrete data collection projects.
He defines the project scope, speaker targets, recruitment approach, and delivery expectations before production starts.
Emil supports operations, documentation, compliance coordination, and project delivery.
He helps structure the process so recruitment, consent, production, and handoff stay aligned.
Gustav leads the platform architecture behind Spirelight.
He builds the systems used to manage recording, transcription, QA, metadata, contributor workflows, and dataset delivery.
Joyi manages project execution across contributors, reviewers, and delivery teams.
She keeps production moving, follows up on daily progress, and helps ensure each project meets its agreed requirements.
Mateo coordinates contributors, recording workflows, and production tasks.
He helps translate project requirements into daily execution and keeps the different parts of the workflow aligned.
Pekka supports recruitment strategy and contributor operations.
He helps source speakers for projects with specific language, dialect, regional, or profile requirements.
Victor leads sales and market research, mapping where Spirelight's speech data work fits new clients and regions.
He runs country expansion research and opens conversations in the markets we move into next.
Yusif leads legal, internal policy, and compliance at Spirelight.
He maintains the internal policies and compliance processes that keep consent, data handling, and contracts in order.
Michael leads B2B sales and manages Spirelight's partnerships with freelancers and companies.
He builds the relationships that bring in new clients and contributors, and keeps our freelancer and company partners aligned with each project's needs.
Practical buyer guides for teams building voice AI: what speech data is, how much a project may need, and which licensing terms to evaluate.
The building blocks of voice AI training data, in plain terms.
02Vendors, licensing, and quality checks before you commit budget.
03Sizing a dataset for the accuracy your model has to hit.
04What makes a speech recognition dataset actually accurate.
05Types, methods, and what to expect from an annotation partner.
06How to define and verify accent, dialect, and low-resource-language coverage.
07What it takes to build a natural text-to-speech voice.
08How emotion in speech is collected, labeled, and used.
09Types, workflows, and evidence to request when labeling audio.
10How Appen, Defined.ai, Shaip, and specialists compare, and how to choose.
11How to verify whether a free dataset's license fits a proposed commercial use.
12What collection can cover and which assumptions to verify in a quote.
13Buyer questions for language, cabin, channel, and evaluation coverage.
14Role-specific data governance and vendor-diligence questions for voice teams.
Services and advanced topics: speech data collection services, audio annotation services, speaker diarization, speech data licensing and license agreements, telephony speech datasets, what speech data collection costs, or what data annotation is.
Popular collection configurations: Russian collection configuration, Amharic collection configuration, Egyptian Arabic collection configuration, Marathi collection configuration, Italian collection configuration, Greek collection configuration, or review all collection configurations.
Looking for paid contributor work? Create a free AI training profile, or read how people make money training AI. Projects open in waves and matching depends on each brief.
Native contributor teams matched to the language and regional variant in the recording, with the written convention agreed before production starts.
Staffed by the recorded variant, from Castilian to Rioplatense.
02Coverage from Bavarian to Low German, convention agreed upfront.
03Neapolitan, Sicilian, and Venetian handled as their own regional languages.
04Seoul, Gyeongsang, and Jeolla audio, speech level preserved.
05Character set, segmentation, and number conventions fixed upfront.
06Script policy across kanji, hiragana, and katakana, agreed upfront.
See how the global contributor workforce behind every transcription project is recruited and managed.
Send over your speaker profiles, language needs, and recording conditions. Our team will review the brief and respond with feasibility questions and next steps.
A project can scope wake-word and command speech around the speakers, accents, devices, and acoustic environments relevant to the intended deployment, subject to recruitment and recording feasibility.
Where this fits: automotive, smart home, appliances, wearables, voice assistants.
When existing data does not cover a target market, the brief can define native-speaker, region, and dialect criteria. The verification method and feasibility are confirmed before commitment.
Where this fits: companies launching voice products in new languages, accents, or dialects.
Emotion and intent annotation can be included when the buyer defines the tag set, reviewer qualifications, context rules, QA method, and acceptance criteria.
Where this fits: voice agents, call centers, automotive safety, health tech, gaming, and any product that needs natural human-machine interaction.
Datarade
