Verbatim ground truth
Word-for-word including false starts, fillers, and non-standard pronunciation. The literal audio, on the page.
VERBATIM · FILLERS · TIMESTAMPS
Transcripts, [verbatim] and on-spec.
Word-level, multi-speaker, multilingual transcripts from verified linguists. Timestamped, diarized, and delivered in the format your pipeline already reads.
Automated transcription is right most of the time. Your model needs it right every time.
Whisper and its peers land at 5 to 15 percent word error, depending on accent, noise, and domain. For human reading, that is fine. For training data, every error compounds. A model trained on a 10 percent WER corpus inherits those errors as correct answers.
Ground truth means ground truth.
Why this holds up
Built for ground truth, not readability
Verbatim. Word-level timestamps. Speaker labels. Phonetic annotations where your pipeline needs them. No cleaned-up reads, no silent edits.
Verified linguists, native speakers
Transcription done by people who grew up speaking the language, not learning it. Accent, idiom, and code-switching captured the way the audio actually sounds.
What the audio actually means
Verbatim captures the words. Emotion capture holds the meaning. Anger, joy, hesitation, sarcasm, stress, pause. Both layers ship with every transcript.
Domain [precision]
Medical, legal, financial, technical. Transcribers are matched to your terminology, not reassigned from a general pool. The jargon lands right the first time.
What we transcribe
Diarized multi-speaker
Speaker-labeled transcripts with turn-level timestamps, including overlapping speech. Built for conversational AI.
DIARIZED · TURNS · OVERLAPPING
Word-level timestamps
Every word time-aligned to the audio, typically within 20ms. Required for ASR training and evaluation pipelines.
ALIGNED · MS-LEVEL · CTM
Phonetic annotation
IPA-level transcription for TTS preparation, pronunciation modeling, and speech research. Prosody markers on request.
IPA · PHONEMES · PROSODY
Emotion & paralinguistic
Tone, emotion, stress, hesitation, laughter, and silence annotated alongside the text. The meaning under the words.
EMOTION · PROSODY · NON-VERBAL
Multilingual & code-switched
Native-speaker transcribers across 60+ countries and 50+ languages. Code-switched audio handled without losing either language.
NATIVE · DIALECT · CODE-SWITCH
How a transcription project runs
Scope
Tell us the audio (format, hours, domain, languages), the output format you need, and the quality threshold. We return a scoped plan and a sample transcript from your actual audio in 48 hours.
Transcribe & annotate
Native-speaker linguists with matched domain expertise transcribe to your spec. Every file is annotated with timestamps, speaker labels, emotion tags, and any additional layers your pipeline requires.
Verify & deliver
Peer review plus centralized QC on every transcript. Delivered in your format: JSON, CTM, TextGrid, SRT, VTT, or a custom schema mapped to your pipeline.
The Human Standard, applied to every transcript.
What ships with every transcript
- Transcript
- Verbatim text, aligned to the audio
- Timestamps
- Word-level, millisecond accuracy
- Speakers
- Diarized with unique speaker IDs
- Emotion
- Tone, prosody, and paralinguistic tags where scoped
- Annotations
- Phonetics and domain layers where scoped
- Linguist
- Verified transcriber ID, anonymized for delivery
- Review log
- Every QC decision, with actor and time
- Format
- Your schema: JSON, CTM, TextGrid, SRT, VTT, custom
- Card
- Per-delivery documentation, covering everything above
Every transcript ships with its receipts.
Who it's for
ASR model teams
Ground-truth training data with word-level alignment, domain coverage, and schema-compatible delivery formats.
TTS and voice synthesis teams
Phonetically accurate transcripts with prosody markers, matched to studio recording sessions for production synthesis.
Conversational AI and voice-agent teams
Diarized multi-speaker transcripts with emotion and turn-taking annotation for dialogue modeling and agent evaluation.
Questions
Tell us about the audio.
Share the audio, the languages, and the output format you need. We come back within one business day with a sample transcript from your actual audio.