Collect new speech data, or turn your existing audio into training-ready transcripts and labels. We scope the speakers, recording conditions, quality checks, and usage rights around your project.
Find the service you needSpirelight is a speech data company. One written scope can cover speaker recruitment, custom recording, transcription, labeling, QA, rights, and delivery. Buyers choose custom work when no existing dataset fits the language, accent, recording conditions, label schema, or acceptance test they need.
Our work splits into two core services. Data collection records new audio to your spec, from scripted prompts to spontaneous, far-field, and in-car speech, in the languages and regional varieties your project needs. Annotation turns raw recordings into labeled training data, with transcripts, timestamps, speaker tags, and events.
Recruitment, consent, recording, annotation, QA, rights, and delivery artifacts are defined for each project. When a dataset that already exists will do, the catalog lists 60 conversational speech datasets with their hours, speakers and price per hour.
The video couldn’t start. Try playing it again or open the video directly.
Order a custom collection, have your audio annotated, or license an off-the-shelf dataset in your language.
Custom recording briefs for scripted, spontaneous, wake-word, far-field, and in-car speech, on-site in controlled settings or remote. Each project defines its target speakers, demographic quotas, consent needs, output structure, and acceptance criteria.
Transcription, speaker and event labels, word-level timestamps, and intent and emotion tags on your audio. The QA workflow and acceptance criteria are agreed for each project.
Sixty conversational speech datasets, one per language or dialect, with dual-channel audio, time-aligned transcripts and speaker metadata. Each page states its hours, speakers and price per hour, and free samples are sent on request.
Each offer is scoped to a specific speech AI application and covers its collection, consent, annotation, and evaluation requirements.
Private held-out test sets built around your callers, channels, languages, accents, noise, and failure modes. References are human-reviewed, and the holdout rules are documented.
Simulated customer-service calls, recorded with consent. Delivered as telephony audio, with separate channels as an option, human-reviewed transcripts, and custom scenarios for your domain.
Separate-channel conversations that keep natural overlap, interruptions, backchannels, repairs, and turn timing. Built for speech-to-speech and low-latency voice systems.
Synced audio and video for visual speech, lip alignment, audiovisual agents, and digital-human research. Participant releases state the intended model use.
Short commands, numbers, item codes, confirmations, and corrections, recorded for the headsets, languages, and noise of your warehouse floor.
Simulated calls made with consent, never real victim calls. They cover multilingual scenarios, urgency, interruptions, and degraded phone audio, with human-reviewed reference transcripts.
Emotion labels for each turn of speech, in discrete classes, with optional valence and arousal ratings. Intent is labeled separately, and annotator agreement is reported per class.
Emotion and sentiment labels for each turn of call audio, for quality monitoring, escalation prediction, and agent assist. The consent chain is documented.
Affect labels for each turn of spontaneous conversation, so a voice agent can detect frustration and be tested on how it responds.
Recruited listener panels rate expressive synthetic speech for mean opinion score, naturalness, and emotional fit. Agreement is reported for each rater.
Custom video recorded to your spec: talking heads and dialogue, expression and gesture sets, liveness sequences, and in-cabin footage. Every participant gives consent.
On-camera monologue and dialogue for avatars, lip sync, and digital humans. Framing is fixed, sync is frame-accurate, and likeness releases name synthetic-media use.
Pairs of genuine and spoofed speech for training deepfake detectors. TTS, voice conversion, and replay attacks are made only from speakers who agreed to be spoofed.
Native contributor teams are matched to the language and regional variant in the recording. The written convention is agreed before production starts. Start from the full transcription service overview, or see how we recruit and manage the global contributor workforce behind every project.
Annotators staffed by the recorded variant, from Castilian to Rioplatense, with the output convention fixed before the first file.
Coverage from Bavarian to Low German, with an orthographic convention agreed before production starts.
Neapolitan, Sicilian, and Venetian handled as regional languages in their own right, not as accented Italian.
Seoul, Gyeongsang, and Jeolla audio, with the speaker's actual speech level preserved throughout.
Mandarin transcription with the character set, segmentation, and number conventions fixed upfront.
Hong Kong audio delivered as written Cantonese or Standard Written Chinese, in traditional characters, with code-mixed English kept.
A written script policy across kanji, hiragana, and katakana, agreed before the first file.
Dialect coverage across Egypt, the Gulf, the Levant, and Morocco, staffed by market.
France and francophone Africa staffed separately, with code-switching handled by convention.
Brazil and Portugal staffed separately, because the two markets do not share one annotator pool.
Sourced across Russian-speaking markets, with a stated convention for mixed contact speech.
Mixed English and Hindi speech transcribed to an agreed script policy, not treated as noise.
Loanword conventions and dialect assignment settled before the first file is processed.
Send your speaker profiles, language needs, and background noise conditions. We will check what is feasible and tell you what we need to scope a written project proposal.
