Transcription
Verbatim or clean-read, to a written convention covering fillers, restarts, repetitions, mispronunciations and inaudible spans.
- Verbatim with disfluencies
- Clean-read for readability
- Inaudible and overlap marking
- Non-speech event tags
Speech & conversation
Japanese speech and audio data annotation
Transcription, timestamping, speaker diarization, intent and slot labeling, and emotion or dialect tagging on Japanese audio — produced against a written convention rather than an unwritten habit.
Two transcribers given the same Japanese audio and no convention will disagree on roughly every third line. Not on what was said, but on how to write it.
The transcript is a set of decisions, not a recording
Japanese has no spaces, four scripts and a wide gap between how a word sounds and how it is conventionally written. So the moment audio becomes text, someone has made dozens of choices: whether 有難うございます is written in kanji or kana, whether 3人 is 3人 or 三人, whether the えーっと at the start of a sentence is kept or dropped, whether a mid-word restart is transcribed or repaired.
None of those choices is wrong in isolation. What is wrong is making them inconsistently across a corpus, because an ASR model trained on inconsistent targets learns that the same sound maps to several strings, and its error rate absorbs the difference. The single largest quality lever in Japanese speech data is a normalization convention that is written down and actually enforced.
So we start every speech project by agreeing that document — and where you already have one, by adopting yours rather than imposing ours.
At a glance
- Task types
- Verbatim and clean-read transcription, timestamping, diarization, intent and slot labeling, emotion and prosody, dialect tagging.
- Audio types
- Call centre recordings, meetings, interviews, broadcast, in-car and device speech, spontaneous conversation.
- Conventions
- Script choice, fillers, numerals, laughter and inaudible spans are all defined before transcription starts.
- Typical output
- JSON with word or segment timings; SRT, VTT, CTM and TextGrid also supported.
- Review
- Second-pass listening on every file, plus blind re-transcription on a sample.
Task types
What we annotate
Transcription is usually the base layer; the rest are added on top of the same timeline.
Timing and alignment
Segment, utterance or word-level timestamps, and forced alignment against an existing transcript where one exists.
- Utterance boundaries
- Word-level timings
- Forced alignment QA
- Silence and overlap spans
Speaker labeling
Who is speaking, across overlapping speech and channel changes — the part that most often breaks on real call recordings.
- Diarization and speaker IDs
- Role labels (agent / customer)
- Overlap handling
- Speaker attribute tagging
Intent and slot labeling
What the utterance is trying to do and what values it carries, layered onto the transcript for voice assistant and call analytics work.
- Intent classification
- Slot and entity spans
- Dialogue act labels
- Call outcome and disposition
Paralinguistic labeling
Emotion, sentiment, prosodic emphasis, speaking style and regional variety, scored against defined categories rather than impressions.
- Emotion and sentiment
- Speaking style and politeness level
- Regional variety and accent
- Audio quality and environment tags
Japanese specifics
The decisions a Japanese convention has to make
These four cover most of the disagreement between two competent transcribers.
Script choice for the same word
表記ゆれありがとう, 有難う and アリガトウ are the same word. Left to each transcriber, a corpus will contain all three, and the model treats them as different targets.
How we handle it The convention lists the required script for common words, sets a default rule for the rest, and an automated check flags variants before delivery.
Numbers and counters
数詞・助数詞Japanese counters change reading with the noun they count — 一本, 一杯, 一人 — and a number can be written as 3, 三 or 参. Written inconsistently, they make numeric accuracy in the resulting model unmeasurable.
How we handle it A numeral policy fixes Arabic versus kanji per context, and how counters and units are transcribed, with worked examples for dates, money, times and phone numbers.
Fillers and disfluencies
フィラーあの, えーっと, まあ and そのー are pervasive in spontaneous Japanese. Keeping them makes transcripts noisier; dropping them makes verbatim data unusable for disfluency-aware models.
How we handle it Verbatim and clean-read variants are defined separately, with a fixed filler inventory and spelling for each, so the choice is a project setting rather than a per-file judgment.
Honorific register in call audio
敬語Call centre Japanese is dense with honorific and humble forms, and several are commonly misheard — いたします and いただきます, よろしいでしょうか and よろしかったでしょうか — by transcribers working at speed.
How we handle it Reviewers are briefed on the honorific patterns specific to your audio type, and honorific errors are a named category in the review pass rather than being folded into general accuracy.
Process
How a speech project runs
The convention document is written before anyone transcribes a full file, and revised only through a versioned change.
-
01
Sample and scope
We listen to a representative sample — including your worst audio, not just your cleanest — and identify what will actually be difficult about it.
-
02
Write the convention
Script, numerals, fillers, non-speech events, overlap, inaudible spans, speaker roles and timing granularity are all fixed in a document you approve.
-
03
Calibrate on a pilot
Several transcribers work the same files independently. Where they diverge, the convention was ambiguous and gets sharpened.
-
04
Transcribe and review
Every file gets a second-pass listen by a different person; a sample is re-transcribed blind to produce a measured accuracy figure.
-
05
Deliver and report
Transcripts, timings and labels ship with the convention version, the QA report and a list of files flagged as unusable rather than guessed at.
Delivery
Formats in and out
We work from your audio as it is. Where a codec or sample rate limits what is achievable, we say so before the project starts rather than after.
| Layer | Default output | Also available |
|---|---|---|
| Transcript | JSON with segment timings | plain text, SRT, VTT |
| Word-level timing | JSON | CTM, TextGrid |
| Speaker labels | RTTM or JSON speaker turns | per-channel transcripts |
| Intent and slots | JSONL, one utterance per line | CSV, CoNLL span format |
| Paralinguistic labels | JSONL with time spans | CSV, ELAN annotation files |
Quality
Controls specific to this work
Accuracy on Japanese audio is measured, not asserted — and measured on the difficult files, not the easy ones.
- Blind re-transcription of a sample by a second transcriber, producing a measured error rate per batch.
- Automated convention checks for script variants, numeral format and filler spelling before delivery.
- Timing validated against the audio, not just against neighbouring segments.
- Files that genuinely cannot be transcribed are flagged and returned as such, never filled in with a plausible guess.
- Speaker labels checked across the whole file, so an identity swap halfway through is caught.
- Difficult audio deliberately included in the QA sample, so the reported figure reflects the corpus rather than its best part.
FAQ
About speech and audio
What teams ask before sending their first batch of Japanese audio.
Yes, and for clean audio it is usually the better value. The caveat is that correcting a machine transcript biases the human toward accepting what is already on screen, so for training data we either transcribe from scratch or run a blind check on a sample to measure how much the correction pass is actually catching.
We agree in advance whether dialect is transcribed as spoken or normalised to standard Japanese, because that decision changes the dataset entirely. Where dialect is in scope we tag the variety, and we are candid about which varieties we can staff reliably and which we cannot.
We work with real recordings, including telephone-bandwidth call audio, background noise and overlapping speakers. What matters more than the format is that you send a representative sample up front — scoping against clean audio and then delivering against noisy audio is how speech projects go wrong.
Yes, under an NDA and a data processing agreement, in an isolated environment with role-based access, and with retention and deletion terms agreed in the contract. Where recordings contain personal information we can also mark it for redaction, or work on audio you have redacted before sending. See the security page for the full picture.
Yes. Overlap is where automatic diarization fails most often, and it is one of the main reasons to put a human on it. Our convention defines how overlapping speech is segmented and attributed, so the treatment is consistent rather than decided file by file.
Speech & audio
Send us your hardest audio, not your cleanest.
A representative sample tells us more than a specification does. We come back with a draft convention, a transcribed sample and an honest view of the error rate to expect.
NDA before you share any data. Pilot scoping is free.