Audio training data · already transcribed · word-level timestamps
50M+ hours of speech from public sources, across 90+ languages and a wide range of accents. Every hour comes with an AI transcript and word-level timestamps, so it is ready to train speech-to-text, text-to-speech and voice agent models as soon as you get it.
nobodynoticeslatencyuntilthemomentitstopsbeingthereandthenit'salltheyhear
Matched audio and text is what speech recognition, speech synthesis and multimodal models train on. The same hours work for all three.
Conversational, accented and noisy audio, including phone calls — closer to what your users actually say than scripted studio recordings.
Natural timing, pauses and emphasis from unscripted speech, with the alignment you need to model how long each word takes.
Each hour is a matched pair of audio and text, which is what speech LLMs, multimodal large language models and voice agents train on.
This is the manifest that comes with every order: one row per audio file, with its word-level timestamps as JSON. Ask for a sample and we’ll send audio and transcripts for the languages you need.
| audio file | duration | rate | language | word-level timestamps |
|---|---|---|---|---|
| ac-0e91f.webm | 11:42 | 48 kHz | en-US | [{"word":"nobody","start":0.18,"end":0.62},{"word":"notices","start":0.62,"end":1.08},{"word":"latency","start":1.11,"end":1.68},…] |
We will send you 50 hours picked at random from the catalog, with their word-level transcripts, at no cost. No call needed.
$0.15 per audio hour at any volume, with transcripts included. Discounts start at 0.5M hours and increase from there.
| List | Standard rate, any volume | $0.150 |
|---|---|---|
| Volume | 0.5M hours and up | $0.135 |
| Scale | 1M hours and up | $0.120 |
| Corpus | 2.5M hours and up | $0.105 |
400,000 more hours unlocks the Volume tier at $0.135/hr.
Request this quote →Estimated pricing. Your final rate is confirmed in your quote and depends on the language mix and any filtering you need.
Tell us what you need and we’ll reply with a price, how many hours are available in your languages, and a free sample.
Prefer email? trainingdata@uksolars.com
We’ll reply by email shortly. If it’s urgent, mail trainingdata@uksolars.com directly.
Publicly available sources. The catalog is speech first — there is little to no music in it. You receive the audio files themselves, with language and duration metadata.
One hour of audio plus its transcript and metadata. Transcripts and word-level timestamps are included in the $0.15 per hour, and so is delivery.
They are AI transcripts with a confidence score on every word, so you can filter before training. Most are verbatim and include punctuation and capitalisation. Request the free sample and measure it against your own reference set.
50 hours selected at random from the catalog, with their transcripts, at no cost.
You get a perpetual, worldwide, non-exclusive licence to train commercial models on it. Redistributing the data itself is not permitted.
No, it is non-exclusive by default, which is what keeps the price at $0.15 per hour. Exclusive arrangements are available on larger orders.
Yes. Tell us the language mix you want and we will tell you how many hours are available.
Upfront.