Zulu speech dataset, roughly 500 hours ready to license
Our OTS Zulu speech dataset is roughly 500 hours of Zulu speech, prompts and conversations, already recorded by Zulu speakers in our 30,000+ network, with consent from every speaker for commercial AI training. License it now and train your Zulu speech recognition model without waiting on a new collection.
- ~500 h of Zulu in stock
- ~35,000 h across the OTS pool
- 100% of speakers consented
- ISO 27001 certified
- SOC 2 report under NDA
- GDPR DPAs and HIPAA BAAs on request
What's in the Zulu speech dataset?
The IndecalAI OTS Zulu speech dataset is roughly 500 hours of Zulu speech, prompts and conversations, recorded by Zulu speakers in our network who all consented to commercial AI training. It's already recorded, so you can license it now instead of waiting on a new collection.
Zulu is the most common home language in South Africa, the continent's most developed market for banking and call centers. Clicks and mixed urban speech are why it needs speech data of its own.
Every recording comes with a transcript and metadata about the speaker, the device and the setting, and you can take it as WAV, OGG or another common format. Every order starts with sample clips and a spec sheet, so you can match the data against your model before you agree a licence.
Need more than what's in stock? Zulu speakers from our 30,000+ network record new speech to your brief, and our Zulu data page covers what we collect to order.
Zulu speech dataset at a glance
- Hours
- Roughly 500 hours, already recorded
- Language
- Zulu (isiZulu), zu
- Content
- Speech, prompts and conversations
- Consent
- Every speaker consented to commercial AI training
- Speakers
- Zulu speakers from our 30,000+ network
- Audio and transcripts
- Transcripts and metadata with every recording, in WAV, OGG and more
- Licence
- Terms agreed with each customer, in writing
- Price
- Quoted per order
- Samples
- Zulu sample clips sent in reply to your request
- Security
- ISO 27001 certified. SOC 2 report available under NDA. GDPR DPAs and HIPAA BAAs on request.
What comes with every Zulu recording
Recorded on the devices and in the places your users actually speak, and delivered in the format your pipeline expects.
| Detail | In the Zulu dataset |
|---|---|
| Speech types | A mix of scripted prompts and natural conversations |
| Recording devices | Mobile phones, laptops and studio microphones |
| Environments | Quiet rooms and noisy, everyday settings |
| Transcripts | Included with every recording |
| Speaker metadata | Speaker details, such as gender and age, for each recording |
| Recording metadata | The device and the setting for each recording |
| Audio formats | WAV, OGG and other common formats |
| Sample rates | Listed for each recording on the spec sheet |
What roughly 500 hours of Zulu speech looks like
Roughly 500 hours sounds abstract until you play it: about 21 days of Zulu speech without a pause. That's why we send sample clips and a spec sheet first, so you can judge the whole from a part.
- ~500 hours of Zulu speech
- ~30,000 minutes
- ~1.8M seconds of audio
- ~21 days, played back to back
Hours here are recorded audio. When you compare quotes, check whether each one counts audio hours or only the parts where someone is talking.
Zulu at a glance, for speech teams
The facts that shape a Zulu speech project, from who speaks it to how it's written. Our Zulu data page covers what we collect to order.
Status
Official language
Official in South Africa, where it's the most common home language.
Where it's spoken
South Africa
Spoken mainly in KwaZulu-Natal and Gauteng, and understood across much of South Africa.
Script
Latin alphabet
Written in Latin letters, with click sounds written as c, q and x.
Family
Bantu
A Nguni Bantu language, close to Xhosa, Swati and Ndebele.
Varieties
Rural and urban
Rural KwaZulu-Natal speech differs from Johannesburg speech, which mixes in English and other languages.
Codes
zu · zul
ISO 639-1 and ISO 639-3 codes. Locales speech APIs often use: zu-ZA.
Zulu speech dataset, voice dataset or speech corpus?
Speech dataset, voice dataset, corpus: the names overlap. Here's how they differ, and where our Zulu data sits.
Zulu speech dataset
The general term
Recorded Zulu speech with details about each recording, for training and testing speech models. The name this page uses.
Zulu voice dataset
Usually the same thing
Some teams mean voice for text-to-speech or speaker recognition, which needs cleaner audio. Part of our data was recorded in a studio, so ask for those recordings.
Zulu audio dataset
A wider term
It can include music or background sound. Ours is speech audio: Zulu speech, prompts and conversations.
Zulu speech corpus
The research name
A structured collection of recordings and their details. Commercial datasets are corpora too, licensed instead of published.
Zulu ASR or speech-to-text dataset
Speech plus what was said
For training a model that turns Zulu audio into text. Every recording in our Zulu dataset comes with a transcript.
isiZulu speech dataset
Same language, other name
isiZulu is what Zulu is called in Zulu. Searches for isiZulu speech data mean this dataset.
What makes Zulu hard for speech recognition
Four things that catch out Zulu speech models, and what to listen for when you check a sample from anyone, us included.
English words inside Zulu
Urban Zulu speakers mix in English all the time, especially for numbers, money and tech. A model trained only on formal Zulu stumbles on that, so listen for it in a sample.
Clicks in the spelling
Zulu writes its click sounds with c, q and x, and joins prefixes to words. Transcripts have to spell clicks and word breaks consistently.
Accents across Zulu speakers
Durban, rural KwaZulu-Natal, Johannesburg and Soweto speech carry different sounds and mixes. Check that the speaker mix covers your users.
Click consonants
Zulu has three click sounds, written c, q and x, which most speech models never hear in training. They need Zulu speech to learn them.
What teams build with a Zulu speech dataset
Where ready-recorded Zulu speech tends to go. These are kinds of work, not client names.
Zulu speech recognition
Train or fine-tune a Zulu speech-to-text model, or adapt a multilingual model so it handles Zulu well.
Voice bots and IVR
Teach a voice bot how Zulu callers actually ask for things, in the words and accents they use.
Customer service calls
Zulu speech for models that route, transcribe or summarize customer calls about accounts, orders and claims.
Voice in the car and at home
A Zulu base for in-car assistants and smart devices, before recording commands in the real setting.
Support and speech analytics
Zulu conversations for models that find intent, sentiment or topics in support calls and chats.
Testing a Zulu model
Held-back Zulu recordings your model never trained on, to measure word and character error rate before launch.
How many hours of Zulu speech do you need?
It depends on the model you start from and how close your users sound to the training data. These are rough starting points that come up often in speech work. Your own tests should decide.
| Goal | Rough starting point | With our Zulu data |
|---|---|---|
| Test a model you already have | A few hours your model never trained on | A test set cut from stock and kept out of training |
| Fine-tune a multilingual model, such as Whisper or wav2vec 2.0 | Tens to a few hundred hours | Roughly 500 hours in stock covers it |
| Adapt to one domain, like banking calls | A smaller set that sounds like your users | Stock as a base, plus a custom top-up in your domain |
| Train a Zulu model from scratch | Thousands of hours | Stock as a base, plus a larger custom collection |
Free Zulu speech datasets or a licensed one?
Public corpora such as the NCHLT speech corpus, Mozilla Common Voice and FLEURS have done a lot for Zulu speech research. They cover what their creators chose to record, like read sentences, crowd recordings or broadcast speech, and each comes with its own licence and terms of use. Here's how they compare with a licensed dataset.
| Topic | Public Zulu corpora | Our Zulu speech dataset |
|---|---|---|
| Licence | Set by each corpus, and some limit redistribution or changes | Agreed with you in writing, for what you're building |
| Consent | Covered by each project's own terms | Every speaker consented to commercial AI training |
| What it sounds like | Whatever the project recorded | Speech, prompts and conversations, with sample clips first |
| Who else trains on it | Anyone who downloads it | Set in the licence we agree with you |
| Filling gaps | Up to you | Recorded to your brief by the same network |
| Support | Community forums | A named contact from sample to delivery |
How to check a Zulu speech sample
Half an hour with the Zulu clips and the spec sheet answers most questions. The list works for any vendor, not only us.
| Check | What to look for | Where to find it |
|---|---|---|
| Accents | Speakers from the regions your users come from | The speaker mix on the spec sheet, and the clips |
| Everyday speech | The way your users really talk, including any mixing with other languages | Listen to the clips |
| Script | Transcripts in the script and spelling your model should output | The transcript format on the spec sheet |
| Audio | A sample rate and recording setup close to your product | The sample rate on the spec sheet |
| Numbers | Amounts, dates and phone numbers written one consistent way | Transcripts of the clips |
| Rights | Consent that covers commercial AI training, and a licence that fits | Ask for the licence terms before you buy |
From Zulu sample clips to licensed data
Your named contact runs it from the first question to delivery.
-
01
Ask
Tell us how many Zulu hours you need and what the model is for.
You get A reply within one business day
-
02
Sample and spec sheet
We send Zulu sample clips and the spec sheet in reply.
You get Zulu clips and the spec sheet
-
03
Licence and quote
We agree the licence terms and send a quote for the order.
You get Written licence terms and a quote
-
04
Delivery
The licensed Zulu speech is delivered to you.
You get Licensed Zulu speech with transcripts
-
05
Fill the gaps
Accents, devices or domains that are missing get recorded to your brief.
You get New Zulu speech, recorded to your brief
Get Zulu sample clips Start with the sample clips, then decide.
Zulu speech data terms
The words that come up when you buy or build Zulu speech data. The full glossary covers every data type.
- Click consonant
- A sound made by sucking air in with the tongue, written c, q and x in Zulu.
- Noun class
- A group of nouns that share prefixes and agreement, as in Zulu umuntu and abantu.
- isiZulu
- The name of the Zulu language in Zulu.
- Speaker metadata
- Details recorded about each speaker, such as gender and age, that let you check the speaker mix against your users.
- OTS dataset
- Short for off-the-shelf dataset: data already collected and ready to license before a buyer asks for it, instead of recorded to one buyer's brief.
- Word error rate (WER)
- The share of words a speech model gets wrong against a reference transcript. Lower is better.
- Character error rate (CER)
- The same idea counted in characters, often used alongside WER, since small spelling differences can make WER look worse than the model is.
- Utterance
- One stretch of speech by one speaker, such as a single prompt read aloud or one turn in a conversation.
More Zulu and speech data
Every language in stock is on the off-the-shelf speech datasets page. Related speech datasets: Xhosa, Swahili and English.
OTS speech datasets
Around 35,000 hours across languages, ready to license.
Zulu data to your brief
Zulu speech, text and translation, collected to order.
Transcription
Zulu audio turned into timed text, checked by people.
Languages
100 languages, collected to order.
Data and model evaluation
Held-back Zulu speech scored against your targets.
Questions about the Zulu speech dataset
Straight answers about licensing Zulu speech. The off-the-shelf FAQ covers questions that apply to every language.
Do you have a Zulu speech dataset ready to license?
Yes. Roughly 500 hours of Zulu speech, prompts and conversations are already recorded and ready to license, as part of our off-the-shelf speech pool.
What's in the Zulu speech dataset?
A mix of scripted prompts and natural conversations in Zulu, recorded on mobile phones, laptops and studio microphones, in quiet and noisy settings. Every recording comes with a transcript and metadata about the speaker, the device and the setting, in WAV, OGG or another common format.
Does it include Zulu-English mixed speech?
Urban Zulu often includes English words, so ask how much mixed speech is in the hours you'd license. The sample clips let you hear it before you commit.
Did every speaker consent to AI training?
Yes. Every speaker gave consent for their recordings to be used for commercial AI training.
How is the Zulu dataset licensed?
Licence terms are agreed with each customer, in writing, before delivery. That includes how you can use the data and whether anyone else can license the same recordings.
How much does the Zulu speech dataset cost?
We quote each order on its own, based on the hours you need and the licence terms. Send your request and your named contact comes back with a quote.
Can we hear Zulu samples first?
Yes. Ask and we send Zulu sample clips with the spec sheet, so you can check the audio before you commit.
What if we need more than 500 hours, or a specific accent?
We collect it. Zulu speakers from our network record new speech to your brief, such as a region, an age group, a device or a domain, and it can sit alongside the hours you license from stock.
Is this the same as a Zulu voice dataset or speech corpus?
For most uses, yes. Speech dataset, voice dataset, audio dataset and speech corpus are names for the same thing: recorded Zulu speech with details about each recording. If you need studio-quality voice for text-to-speech, ask for the recordings made in a studio.
Can we use it as a Zulu speech-to-text (ASR) dataset?
Yes. Every recording comes with a transcript, so it's ready for training and testing Zulu speech-to-text models. The spec sheet shows the transcript style, so you can check it matches your pipeline before you license it.
Why not use a free Zulu speech dataset?
Public Zulu corpora are a good start for research. Each one has its own licence and terms of use, and they cover what their creators chose to record, such as read sentences or broadcast speech. Read the licence before you train a commercial model on one, and check that it sounds like your users.
What does OTS mean?
OTS stands for off-the-shelf: data that's already recorded and ready to license, instead of collected after you order. Our OTS Zulu speech dataset is roughly 500 hours.
How many hours of Zulu speech do we need?
It depends on your starting model. A test set can be a few hours, fine-tuning a multilingual model often starts at tens to a few hundred hours, and training from scratch takes thousands. Your own tests should decide.
Can we fine-tune Whisper or wav2vec 2.0 on it?
Yes. Teams commonly fine-tune open speech models like these on licensed Zulu speech, and every recording comes with a transcript in WAV, OGG or another common format. The spec sheet shows the transcript style, so you can plan any conversion before you train.
Do you work with US teams?
Yes. We take calls in US Eastern mornings, which is evening in India. We're ISO 27001 certified, our SOC 2 report is available under NDA, and we sign GDPR data processing agreements and HIPAA business associate agreements on request.
Get Zulu sample clips
Say roughly how many Zulu hours you need and what the model is for. We'll send sample clips and the spec sheet.
- Your named contact replies within one business day.
- We send Zulu sample clips and the spec sheet.
- We agree the licence terms and send a quote.
This opens your email app with the request filled in. Nothing is sent until you press send. You can also use the contact page.