Scripted speech data, recorded from your prompt list
When your model has to recognize specific words, such as product names, addresses, numbers, codes or commands, conversation rarely says them often enough. We record them read aloud by native speakers recruited to your quotas, on the devices and in the environments you specify, then check and transcribe every file in your convention.
Your words, in the voices of your users
A script for speech recognition needs many voices per word. A script for a TTS voice needs one chosen speaker, as in TTS data collection, and a wake phrase needs the phrases that sound like it, as in wake word data collection.
Your prompt list
You send the sentences or prompts: product and brand names, people and place names, addresses, numbers, dates, amounts, codes, domain terms or commands. Speakers read them on the devices and in the environments your brief specifies.
Many voices per word
Native speakers are recruited to your quotas for age, gender, region and accent, so each word is said by the range of people who will say it to your product. Each passes an ID check and a microphone test, and we confirm their language or dialect.
Transcribed as spoken
Files pass automatic checks (SNR, clipping, duration) and a human review, and sessions that miss spec are re-recorded. Transcripts are time-aligned in your convention: clean-read, or verbatim with misreads kept and disfluencies tagged.
What a project can include
| Prompts | Your sentence or prompt lists: names, addresses, numbers, dates, amounts, codes, domain terms, commands |
|---|---|
| Speakers | Native speakers recruited to your quotas for age, gender, region and accent |
| Vetting | ID check and microphone test; language or dialect confirmed |
| Devices | Phone, headset, far-field or in-car, as your brief specifies |
| Transcripts | Time-aligned, human-verified, in your convention, clean-read or verbatim, with disfluency tagging as required |
| QA | Automatic checks plus human review; sessions that miss spec are re-recorded, not shipped |
| Consent | Explicit commercial-use consent from every speaker, documented |
| Delivery | Audio, transcripts and per-speaker metadata in the layout your pipeline expects |
| Languages | Any of our 60 catalog languages and dialects, and others after a feasibility check |
Built for
- ASR for product, brand and place names
- Addresses, numbers, dates and amounts
- Spelled codes and reference numbers
- Domain terms in your field
- Test sets for the words your model gets wrong
Scripted or conversation?
Scripted speech is the right tool when a model must hear particular words from many voices. Read speech is also easier to recognize than conversation: in OpenAI’s Whisper paper, the large-v2 model had a 2.7% word error rate on LibriSpeech’s read audiobooks and 17.6% on CallHome telephone conversations. So a common approach is to train on conversation and add a scripted set for the words it lacks, as our comparison of conversational and scripted speech data explains.
Our off-the-shelf datasets are spontaneous two-person conversations recorded in-country, not read scripts, in 60 languages and dialects, so scripted speech is always recorded to your spec.
Frequently asked questions
What is scripted speech data?
Scripted speech data is audio of people reading prepared text aloud, such as sentences, prompts or word lists, with transcripts aligned to what they said. It is the right tool when a model has to recognize particular words, such as product names, addresses, numbers or commands, from many different speakers. Spontaneous conversation is better for teaching a model how people actually talk.
Where can I get a scripted speech dataset in my language?
SpeechData.ai records scripted speech to order: native speakers recruited to your quotas for age, gender, region and accent read your sentence or prompt lists on the devices you specify. Its off-the-shelf catalog is spontaneous conversation only, so scripted data is always recorded to your spec. You get a feasibility assessment and a quote within one business day of sending your brief.
Can transcripts keep misreads and hesitations?
Yes. Transcripts are time-aligned and human-verified in your convention. A verbatim transcript records what each speaker actually said, including misread words and hesitations, with disfluencies tagged as your convention requires, while a clean-read transcript leaves the disfluencies out.
Which languages can you record?
Any of the 60 languages and dialects in our catalog, and other languages on request. We confirm recruitment feasibility for each language before quoting.
How is scripted speech collection priced?
Per project, based on the language, speaker count and quotas, speech type, devices and environments, transcription depth and timeline. Off-the-shelf datasets have published prices per hour, while scripted collection is quoted once we have seen your brief.
Tell us which words your model needs
Share your prompt list or a sample of it, languages and accents, speakers per prompt, devices and transcript convention. We’ll come back within one business day with a feasibility assessment and a quote.