Service explainers
What is asr training data and how does it work?
Updated 2026-08-01 · 4 min read
Short answer
Transcribed speech corpora built to train and evaluate automatic speech recognition, with verbatim transcription, timestamps and per-token language tagging where code-mixing occurs. In practice the work runs as style guide authored per language, covering numerals, loanwords, script and disfluency rules, then transcriber calibration round with inter-annotator agreement measurement, then first-pass transcription, and you receive audio plus aligned transcripts (json/tsv, or your schema), style guide as delivered documentation. Transcription adds roughly 1-2 weeks per 100 hours after recording, per language.
Key takeaways
- Agreement is measured, not assumed. We report word-level agreement on a held-out sample so you can judge label quality before training on it.
- Typical buyers: ASR training, ASR fine-tuning, Domain adaptation.
- Recruitment approach: Transcribers are native speakers of the target variety, not of a related standard language.
What the service covers
Transcribed speech corpora built to train and evaluate automatic speech recognition, with verbatim transcription, timestamps and per-token language tagging where code-mixing occurs.
- Audio plus aligned transcripts (JSON/TSV, or your schema)
- Style guide as delivered documentation
- Inter-annotator agreement report
- Speaker-disjoint train/dev/test splits
Technical specification
These are defaults, not limits. Where your pipeline requires different values, they replace ours in the statement of work rather than being converted after delivery.
| Parameter | Standard |
|---|---|
| Audio | 16 kHz or 48 kHz PCM WAV |
| Transcription | Verbatim, including disfluencies, false starts and fillers |
| Timestamps | Utterance level by default; word level on request |
| Tagging | Noise, overlap, unintelligible, foreign-language and code-switch tags |
| Normalisation | Raw and normalised text columns delivered separately |
| Split | Train/dev/test splits with no speaker leakage across splits |
How the work runs
- Style guide authored per language, covering numerals, loanwords, script and disfluency rules
- Transcriber calibration round with inter-annotator agreement measurement
- First-pass transcription
- Second-pass native review
- Automated consistency checks against the style guide
- Split generation with speaker-disjoint partitions
Quality control and acceptance
Agreement is measured, not assumed. We report word-level agreement on a held-out sample so you can judge label quality before training on it.
Failures are remedied by re-collection rather than by editing delivered files, because repaired audio carries artefacts that survive into the trained model.
Who this is for
Recruitment for this service works as follows. Transcribers are native speakers of the target variety, not of a related standard language.
- ASR training
- ASR fine-tuning
- Domain adaptation
- WER benchmarking
Timelines
Transcription adds roughly 1-2 weeks per 100 hours after recording, per language.
Multi-language programmes run in parallel rather than in sequence, so a five-language scope does not take five times as long.
Frequently asked questions
What is included in asr training data?
Audio plus aligned transcripts (JSON/TSV, or your schema), Style guide as delivered documentation, Inter-annotator agreement report, delivered against a written specification with acceptance criteria attached.
How long does asr training data take?
Transcription adds roughly 1-2 weeks per 100 hours after recording, per language.
How is quality measured?
Agreement is measured, not assumed. We report word-level agreement on a held-out sample so you can judge label quality before training on it.
Which languages are supported?
Fifteen Indian languages plus Indian English, including Hindi, Marathi, Tamil, Telugu, Kannada.
Related reading
Turn this into a dataset specification
Tell us the languages, speaker count and minutes. You get a written scope, a protocol and a fixed price within one working day.