Speech transcription and ASR data
Morvio delivers transcripts of Arabic and its dialects, written by native speakers and formatted to your specification, for training and testing speech recognition (ASR) models. European and Asian languages on demand.
Transcript: three speakers in a recorded meeting, with timestamps, speaker tags, overlapping speech marked and laughter tagged.
This one was cooler. I got a whole table and everything.
Now I have all about circuits and chips and transponders and— I wrote it all down, because I thought it would be relevant, like all the different parts of the inside of a remote control, but then they're like, you don't actually need this— you [overlap] just need to talk about the case. [/overlap]
[overlap] That's like mine it was like, would you [/overlap] prefer an LCD screen or a multifunction remote control? And then it didn't have like any kind of table, like awesome, I'm glad they asked that question and didn't report the result [laughter].
It's [overlap] really interesting [/overlap] though.
[overlap] I [/overlap] I thought it was a little tricky having to hand around this thing.
Tags: [overlap] … [/overlap], [laughter].
- Price basis
- Per audio hour
- After the pilot, we confirm the production price in writing.
- Who does the work
- Native speakers
- Transcribers are native speakers of the dialect in the audio.
- Review
- Dedicated QA reviewer
- Also a native speaker; each delivery comes with a QA report.
- How to start
- Send a project brief
- Describe the task in five short steps. You do not need to send any data.
Task types
- Verbatim
- every word, filler, false start and repetition, as spoken
- Clean
- readable text with fillers removed and meaning unchanged
- Segmentation and timestamps
- utterance level; word level
- Speaker labels
- who speaks when, including overlap
- Code-switch marking
- each switch between Arabic and another language, tagged to your convention
- Non-speech tags
- noise, music, laughter, silence, unintelligible speech
- ASR correction
- review and correction of your system's transcripts
- Dataset checks
- clip validation, audio–transcript alignment, metadata (dialect, speaker, recording conditions)
Sample
Waveform of the Fig. 01 clip, with three speaker lanes, each speaker's turns, and the overlapping stretches marked.
{
"clip_id": "clip-0001",
"language": "en",
"segments": [
{
"start": 1963.02,
"end": 1974.17,
"speaker": "SPK-1","text": "[overlap] That's like mine it was like, would you [/overlap] prefer an LCD screen or a multifunction remote control? And then it didn't have like any kind of table, like awesome, I'm glad they asked that question and didn't report the result [laughter]." }
]
}[overlap] That's like mine it was like, would you [/overlap] prefer an LCD screen or a multifunction remote control? And then it didn't have like any kind of table, like awesome, I'm glad they asked that question and didn't report the result [laughter].
Tags: [overlap] … [/overlap], [laughter].
Indicative price
Per audio hour.
The price depends on:
- audio quality;
- the number of speakers and how much they overlap;
- dialect;
- verbatim or clean style;
- timestamp detail;
- review depth.
After the pilot, we confirm the production price in writing.
Get an estimateOther languages
Need another language? We recruit and vet native speakers of European and Asian languages for your project. We confirm a language only once vetted speakers are in place, and you see a calibration batch in that language before any volume.
Specification
- Languages
- Arabic (Modern Standard and dialects). European and Asian languages on demand
- Arabic dialects
- Confirmed per project, with a calibration batch before any volume
- Transcripts out
- JSON, TextGrid, SRT, or your schema
- Quality metric
- WER on a reviewed sample
- Conventions
- Your guideline, or our house conventions
- Tools
- Your tool, or a widely used open-source one
How we check it
- Transcribers are native speakers of the dialect in the audio.
- The team leader settles disagreements.
- WER is measured against the reviewed reference. Timestamps, speaker labels and switch tags are checked separately, because a transcript can have the right words and the wrong speaker.
- Guideline questions are logged. Each answer becomes a guideline update for every transcriber on the project.
- Each delivery comes with a QA report.
What our QA report contains
- sample size
- results against agreed targets
- error types
- adjudication notes
A sample batch gives you a real report on your own data.
Request a sample batchQuestions
Do you keep mixed-language speech in one transcript?
Yes. Each language stays in its own script and every switch is tagged, unless your guideline says otherwise.
How do you spell dialect words with no standard spelling?
We agree a spelling list with you during calibration and keep it current, so the same word is spelled the same way in every file.
Can you correct our ASR output instead of transcribing from scratch?
Yes, priced separately.
What audio quality do you need?
Most audio can be transcribed. If you want, we listen to a short sample under NDA and agree how to tag speech that cannot be made out.
Can you also record the audio?
Yes. See Data collection.
Discuss your project
Two ways to start. Both reach Khaled Zerouki.
Book a call. Pick a time in your own time zone.
Book a callRequest a sample batch. A batch of your own data, returned with a QA report, on terms agreed before work starts.
Request a sample batchWant a price first?
Get an estimateWe reply within 1 business day.¹ NDA on request. We can plan and price your project without seeing your data.
¹ Business days are Monday to Friday, UK time, excluding public holidays in England and Wales.