Script or spontaneous speech?
Scripted speech means a speaker reads a fixed prompt, so you control exactly what gets said and can collect the same sentence from many speakers for comparison. Spontaneous speech means a speaker talks freely about a topic or answers a question in their own words, so you get natural pauses, filler words and sentence structure a script would never produce. Most real products need some of both: scripted prompts to guarantee coverage of specific words or phrases, and spontaneous speech so the model sounds like how people actually talk.
We collect scripted prompts in Prompt Audio and natural conversation in Conversation Audio, both built in-house, so a brief can mix the two without switching vendors.
Who should be speaking
Age, gender, accent and region all change how speech sounds, and a model trained on one narrow group will struggle with everyone else. Decide early which groups your product actually needs to work for, not just which group is easiest to recruit.
Our network is 30,000+ people worldwide, with deep coverage of Indian languages and accents well past Hindi and English. Tell us the group you need and we match it, rather than collecting from whoever is already signed up.
Recording environment
A model trained only on quiet, close-mic recordings can fail badly the moment it meets a noisy call center or a phone held at arm's length. Decide what real-world conditions your product needs to handle: background noise, distance from the microphone, indoor versus outdoor, and the actual devices people will use.
We record in the conditions you specify, and every batch is reviewed before delivery so an unusable clip doesn't reach you.
Designing prompts for a scripted speech dataset
A prompt list should cover every word, phrase and sentence pattern your model needs, without being so long that speakers get tired and their delivery goes flat. Randomizing prompt order across speakers, and mixing sentence lengths, keeps the dataset from sounding uniform.
We help write the prompt set as part of the brief, so coverage gets checked before recording starts, not after.
Metadata you'll want later
Speaker ID, age band, gender, accent or region, recording device and background noise level all turn a folder of audio files into a dataset you can actually analyze. Decide what to capture before recording starts. Adding it afterward means re-contacting every speaker, if that's even possible.
Metadata is agreed as part of the brief and delivered alongside the audio, in the format you specify.
How much is enough
There's no single number that works for every project. It depends on how many groups you need covered, how different your use case is from anything the model has already seen, and how the model performs once you test it. A held-back test set, kept separate from anything used in training, is the only reliable way to know if you have enough.
Tell us the goal and we'll help size the collection against it, rather than quoting a number before we know what you need.
A speech data collection brief template
Here's everything above as one brief. Copy it into a document and fill in what you know. Blank lines are fine: they're the questions to settle with your vendor before anyone records a word.
SPEECH DATA COLLECTION BRIEF 1. Purpose What the data is for: (e.g. speech recognition for a banking voice bot) Training, testing, or both: 2. Speech type Scripted (read) share: __ % Spontaneous share: __ % Prompt set: we supply it / vendor helps write it Topics for spontaneous speech: 3. Speakers Languages and accents/regions: Age bands: Gender split: Speakers per group: Most audio any one speaker may contribute: 4. Recording conditions Places: (quiet room, office, car, street) Devices: (phone, laptop, headset) Channel: (close mic, phone line, both sides of a call) Audio format: (sample rate, bit depth, file type) 5. Volume Total hours or utterances: Held-back test set: (kept apart from training) 6. Transcription Style: verbatim / clean Timestamps: per line / segment / word Tags: noise, laughter, silence, unclear speech Spelling and script rules: (e.g. for mixed-language speech) 7. Metadata with every clip Speaker ID, language, accent or region, age band, gender, place, device, background noise level 8. Checks before delivery What makes a clip unusable: (clipping, cut-off words, wrong speaker, too noisy) Sample to review before full volume: 9. Delivery File naming and folder layout: Metadata format: (CSV, JSON) Batch schedule: Data handling: (NDA, who can access it, retention, deletion)
These are the same questions we ask on every speech project, and the NDA is signed before any data moves. You can send a filled-in brief, or a half-empty one, through our custom speech data collection page and we'll record a small sample free before you commit.
Warning signs
Be cautious of a vendor who can't say how they recruit speakers, won't record in your real-world conditions, or delivers audio with no metadata attached at all. Those gaps tend to surface only after the data is already unusable.