Audio and speech annotation for ASR and voice AI

Transcription, timestamps, speaker labels and sound events, written to one agreed style and checked against reference files before delivery.

Start with a pilot batch See task types

call_0193.wav · 2 speakers · hi-en

00:00.4SPK_1Hi, I'm calling about my order.00:02.1SPK_2Sure, can you share the order ID?00:04.6[door slam]00:05.3SPK_1Haan, it's 4821. Abhi tak deliver nahi hua.

Audio annotation types

From plain transcripts to speaker-level, time-aligned labels. Pick what your model needs.

Transcription

Speech to text in the style you choose: clean, full verbatim or normalised.

Used for: ASR training, captions, call analytics

Timestamps and segmentation

Start and end times for every utterance, word or phrase.

Used for: Alignment, subtitles, search in audio

Speaker diarization

Who spoke when, with consistent speaker labels through overlaps and interruptions.

Used for: Meetings, call centres, interviews

Audio event tagging

Non-speech sounds such as music, silence, alarms, door slams and background noise.

Used for: Sound detection, smart devices, QA

Emotion and intent

How something was said, not only what: frustration, urgency, calm, and the caller's goal.

Used for: Contact centres, voice bots

Pronunciation and accent

Mispronunciations, accent and dialect labels, and phonetic notes where needed.

Used for: Language learning, accent-robust ASR

Wake words and commands

Labels for short commands and wake phrases, including false triggers.

Used for: Voice assistants, in-car, IoT

Speech quality rating

Human ratings for naturalness and clarity of synthetic or recorded speech.

Used for: TTS evaluation, codec testing

Pick a transcription style before you start

Most transcript disagreements come from style, not hearing. We fix the style in the guidelines first.

Style	What it keeps	Example	Best for
Clean verbatim	Words that carry meaning. Removes fillers and false starts.	"I want to cancel my order."	Analytics, summaries, most ASR
Full verbatim	Everything said: fillers, repeats, stutters, laughter.	"I, um, I want to cancel my, my order [laughs]."	Conversational ASR, speech research
Normalised	Numbers, dates and units written in a standard form.	"Pay 2,499 rupees by 14 March."	Search, entity extraction, TTS
Code-switched	Each language written as spoken, in the agreed script.	"Haan, it's 4821. Abhi tak deliver nahi hua."	Indian-language and Hinglish products

How a project runs

Pilot first, scale only after you have heard real output.

Share samples and goal
A few hours of audio, your language list and what the model needs to do.
Write the guidelines
Transcription style, tags, speaker rules and edge cases. You sign them off.
Pilot batch and QA report
A small batch, partly double-transcribed and scored against your gold files.
Refine
You review. We update rules and fix what you flag.
Scale
Full delivery in batches, each with the same QA report.

What the QA report shows

Acceptance criteria, such as a target word error rate, are agreed before the pilot.

Word error rate against goldSubstitutions, deletions and insertions compared with your reference transcripts.
Timestamp accuracyHow far segment boundaries are from the reference, within the tolerance you set.
Speaker label accuracyCorrect speaker changes and consistent IDs through the file.
Tag consistencyEvent and emotion tags checked against the guideline definitions.
Batch summaryHours delivered, reviewed and reworked.

Tools and formats

We work in your tool or set one up, and deliver in the format your pipeline reads.

Tools

Label StudioELANPraatYour in-house tool

Export formats

JSON with timestampsSRT / WebVTTRTTM (diarization)Praat TextGridCSVPlain text

Data security

Voice recordings are personal data. We treat them that way.

NDA first. Signed before any audio or guidelines are shared.
Personal details handled by rule. We agree upfront how to mark or redact names, numbers and addresses in transcripts.
Need-to-know access. Only transcribers and reviewers on your project can play your files.
No downloads where you say so. Work can happen inside your platform with streaming only.
Used only for your project. Never shared, sold or reused, and deleted on request.

Where teams use it

The task types above, applied to real products.

Contact centres

Transcripts, speakers, intent and emotion on customer calls.

Voice assistants

Commands, wake words, slots and false triggers.

Meetings and media

Diarized transcripts, captions and subtitles.

Indian-language ASR

Hindi, Hinglish and regional-language speech, with code-switching rules.

Text to speech

Naturalness and clarity ratings, pronunciation checks.

Smart devices

Sound event labels for alarms, glass break, appliances and noise.

Questions teams ask

If yours isn't here, ask in the form below.

How is audio annotation priced?

Usually per audio hour, depending on audio quality, number of speakers, transcription style and the tags you need. We quote after hearing samples and confirm the rate after the pilot.

Which languages and accents do you support?

Tell us the languages, accents and scripts you need. We confirm transcriber availability for each before the pilot, rather than promising it upfront.

How do you handle Hinglish and code-switching?

The guidelines fix the script for each language and how to mark switches, so every transcriber writes mixed speech the same way.

What about poor audio quality?

Unclear words get a standard tag instead of a guess, and we report how much of each batch was unclear.

Do you also collect speech data?

Yes. See data collection for scripted and spontaneous speech recording with consent.

Do you also do other data types?

Yes: image & video, text & nlp, llm evaluation, data collection.

Start with a pilot batch

Tell us about your audio and the model you are training. We'll reply with questions and a pilot plan.

Prefer email? Write to hello@norbeon.com.

Run an annotation or data team? Register as a vendor.