Read (scripted) speech
People record prompts you write, so every command, word and number your model needs shows up often enough.
For voice commands and speech recognition
Our tool: Prompt Audio Free sample of read speech
Speech data collection services
Custom speech data collection, transcribed and labeled for AI
Speech data for AI is recorded audio paired with what was said, who said it and where. We collect it to your brief from people in our 30,000+ network, in languages worldwide and deep across Indian languages, and our own tools carry it through transcription, labeling and checks.
Need speech now? We also license around 35,000 hours of off-the-shelf speech datasets, already recorded.
clip 0418 · recorded
recordedtranscript
- 00:00.3 S1 Set an alarm for six thirty.
- 00:02.1 S2 For tomorrow?
- 00:03.0 S1 Yes, [noise] tomorrow.
labels and speaker details
intent: set alarm Indian English car, phone mic 2 speakers
labeled✓ reviewed by a person clip 0418 · ready to deliver
Prompt Audio recording prompt 23 of 40
The short answer
What is speech data?
Speech data is recorded human speech, paired with a transcript and details about the recording, that a speech model uses to learn to recognize or understand spoken language. The audio teaches the sound, and the transcript and labels tell the model what it means.
A useful speech dataset is planned, not scraped. It covers the speakers your product will hear, by language, accent, age and gender, in the places they'll talk, like a quiet room, a car or a busy street.
Most of it comes in two forms. In read speech, people record prompts you wrote. In spontaneous speech, they talk naturally, the way they would on a real call. Speech recognition usually needs both.
Speech data at IndecalAI
- Kinds
- Read prompts, spontaneous speech, conversations and phone calls.
- Coverage
- Speaker mix by language, accent, age and gender, planned with you.
- People
- Speakers from a network of 30,000+ people worldwide.
- Languages
- Languages worldwide, with deep coverage of Indian languages well past Hindi and English.
- Tools
- Our own Prompt Audio, Conversation Audio and Transcription tools.
- After
- Transcription, audio labeling and evaluation by the same team.
- Your data
- NDA first, used only for the agreed scope, returned or deleted when you ask.
- Contact
- One named contact who replies within one business day.
- First step
- A small free sample before you commit.
- Company
- Based in Ahmedabad, India, since 2021, with 150+ projects delivered.
AI showcase
Open up one clip
Here's a single four-second recording, taken apart. Tap through the layers to see what travels with the audio. It's a made-up example, and what your files carry is up to you.
The audio
The recording itself, made the way your product will hear people: on the devices and in the places your users actually talk.
- Devices: Phones, laptops or headsets, as your brief says.
- Places: Quiet rooms, cars, streets or offices.
- Checks: Cut-off words and unusable clips are caught before delivery.
The transcript
What was said, written down by people who know the language and the accent, then reviewed. Our transcription page goes into the options, and our guide on transcription for AI vs for people covers verbatim against clean in more depth.
- Style: Verbatim or clean, whichever your model needs.
- Timing: Timestamps per line, segment or word.
- Tags: Noise, laughter, silence and unclear stretches, marked the same way every time.
The labels
Speaker turns, sound events and, where the task needs it, intents and entities, labeled by people and checked. See data annotation for the full range.
- Speakers: Who is talking, and when each turn starts and ends.
- Sounds: Speech, music, noise and silence, stretch by stretch.
- Meaning: Intents and entities, when your model needs them.
Speaker details
Who spoke and where, recorded with each clip, so you can check coverage and split results by group later.
- Groups: Language, accent, age group and gender.
- Setting: The place and the device.
- Later: Results split by group in evaluation.
What we record
The speech we record
Read prompts and spontaneous talk, phone calls and real-world conditions, planned around the accents and languages your users speak. People record in our own Prompt Audio and Conversation Audio tools.
Spontaneous speech
People talk naturally about a task, with the hesitation, fillers and corrections that scripts leave out.
For models that have to cope with real talk
Our tool: Conversation Audio Free sample of spontaneous speech
Phone and call audio
Two-sided calls at phone quality, for products that hear people down a line. Our conversational AI data page covers the domains.
For call centers, IVR and voice bots
Our tool: Conversation Audio Free sample of phone and call audio
Accents and dialects
The speaker mix planned from the brief onward, around the markets you serve and where your users live.
For speech recognition that works for every user
Real-world conditions
Recordings made in the places your product will be used, from a quiet room to a noisy street, on the devices people really hold.
For assistants that work outside a quiet office
Mixed-language speech
Speech that switches languages inside one sentence, common in India, recorded and written down by an agreed rule.
For multilingual voice assistants
Where it's used
What teams build with it
A few of the jobs speech data ends up doing. These are typical projects, not a client list. Plenty of projects mix data types, so there are pages for image data, video data and text data too. Two example categories: Hindi BFSI call center speech and healthcare conversation speech.
Speech recognition
Audio paired with checked transcripts, so a model learns to turn speech into text for your users' voices.
Voice assistants
Commands and questions said many ways, in the rooms and cars where people use them.
Voice bots and IVR
Phone calls with both sides captured, so a bot learns how callers really talk.
Call analytics
Two-sided calls transcribed with speaker labels, so a model can follow who said what.
Multilingual users
Speech that mixes languages, recorded from the people who actually talk that way.
Testing across accents
Held-back recordings from every accent group, to check your model works for all of them.
Speech data categories
Speech data by industry and language
Each page covers what that kind of speech project usually needs, and each one is built to your brief. Every category, including image, video and text, is on the datasets hub.
Hindi BFSI call center speech
Tamil BFSI call center speech
Spanish BFSI call center speech
Japanese BFSI call center speech
German customer support calls
Arabic customer support calls
Mandarin customer support calls
French retail support calls
Portuguese retail support calls
Hindi IVR and voice bot speech
Spanish IVR and voice bot speech
Indonesian IVR and voice bot speech
Healthcare conversation speech
Tamil healthcare conversations
Hindi in-car voice data
Hindi chat and voice support
Multimodal support chat and voice
Before recording starts
Settle four things before anyone hits record
Size gets all the attention, but these four choices matter more. They're cheap to change now and expensive once the recordings exist. The speech dataset specification guide goes through each one in more depth and ends with a brief template you can copy.
The speaker mix
Who your product will hear, by language, accent, age group and gender, and roughly how many speakers of each.
Read or spontaneous
How much comes from prompts you write, and how much from natural talk. Most speech models need both.
Devices and places
What people record on and where, so the audio sounds like your product will hear it.
Transcription rules
How the audio gets written down: style, timing, tags and spelling, agreed once for the whole set.
How it works
From the first recording to data your model can use
You deal with one named person the whole way. Each step below links to the service that does it.
Brief
You tell us the languages, the speakers and what the data is for.
NDA first
The NDA is signed before any data moves.
Collection
People record in Prompt Audio or Conversation Audio.
Transcription
People who know the language write down what was said.
Labeling
Speaker turns, sounds and meaning get labeled and checked.
Evaluation
Held-back data checks your model across groups.
Languages
Voices from all over the world
We record speech in languages from all over the world and go deep on Indian ones, well past Hindi and English. Our page on Indian language data has the details.
Glossary
Speech project terms, decoded
Terms you'll hear on a speech project, and what they actually mean.
- Speech recognition (ASR)
- Software that turns spoken words into text. It learns from audio paired with checked transcripts.
- Read speech
- Speech recorded from prompts, also called scripted speech, so the words are known in advance and every one you need gets covered.
- Spontaneous speech
- Natural, unscripted talk, with the pauses, fillers and corrections of real conversation.
- Utterance
- One stretch of speech from one speaker, usually a sentence or a short request.
- Speaker metadata
- Details recorded with each clip, like language, accent, age group, place and device.
- Signal-to-noise ratio
- How loud the speech is compared with the background. Noisy audio has a low ratio.
- Code-switching
- Moving between languages inside one conversation or sentence, common in everyday speech in India.
- Word error rate (WER)
- The share of words a speech model gets wrong against a checked transcript. Lower is better.
Related work
Where a speech project goes next
Most speech projects touch several of these. You can also compare all eight.
Data collection
Speech recorded from people in our network.
Transcription
Audio turned into reviewed text.
Conversational AI data
Dialogue and calls for voice products.
Data annotation and labeling
Speakers, sounds and meaning labeled.
Data and model evaluation
Your speech model checked across groups.
FAQ
Questions about speech data
Quick answers to what people ask us about speech data. The services FAQ covers how the services fit together.
What is speech data?
Recorded human speech, paired with a transcript and details about the recording, that a speech model uses to learn to recognize or understand spoken language.
What is the difference between read and spontaneous speech?
In read speech, people record prompts you wrote, so every word and command you need shows up. In spontaneous speech, they talk naturally, with the pauses, fillers and corrections of real conversation. Speech recognition usually needs both.
Do you have ready-made speech datasets, or only custom collection?
Both. We hold around 35,000 hours of off-the-shelf speech, prompts and conversations already collected across our network, and we check that pool first. If it doesn't fit your languages, speakers or recording conditions, we run a custom speech data collection to your brief.
How much speech data do we need?
It depends on the task, the languages and how varied your users are, so there's no standard number. We plan the speaker mix and volume with you, and the free sample lets you check quality before you commit to volume.
Which languages and accents do you cover?
Languages from all over the world, with deep coverage of Indian languages, well past Hindi and English. Tell us where your users are and we plan the accent and dialect mix around that. Our Indian language data page has more.
Can you record in real-world conditions?
Yes. Tell us the devices and places your product will hear people in, like a phone in a car or a laptop in an office, and we plan the recordings around them.
Can you transcribe and label the recordings too?
Yes. The same team handles transcription and audio labeling, so the recordings move on without a second vendor.
Can we try a small sample first?
Yes. Send a short brief and we'll record a small sample free before you commit.
Does our data stay private?
Yes. The NDA is signed before any data moves, we use it only for the scope we agreed, and only the people working on your project can get to it. It's returned or deleted when you ask.
Start with a free sample
Pick the kinds of speech, add the languages and roughly how many hours you need, and send it. We'll record a small sample free before you commit.
- Your named contact replies within one business day.
- We sign an NDA before any data changes hands.
- We record your free sample, then agree the full scope.
This opens your email app with the brief filled in. Nothing is sent until you press send. You can also use the contact page.