Seven stages. You only see the last one.
Every order runs the same pipeline. Four of those stages call an AI model; the other three are ours, and they are the reason the result sounds like radio.
A language model writes imaging copy from your brief, your format preset and your sound kit — in the target language, not translated into it. Call letters are marked for pronunciation, legal wording is locked and never rewritten, and the result comes back structured: lines, emphasis, pauses, target seconds and a suggested layer recipe.
Several takes per line are rendered with an imaging-grade voice, with your pronunciation dictionary applied so local place names and host names land correctly. Each phrase is kept as its own stem, so the mixer can space and punch them individually rather than being handed one flat read.
Instrumental beds are generated keyed and tempo-locked to your sonic logo. For jingles, your station name is sung over your locked melody — rendered once, stored in your kit, and reused over fresh beds so the hook never drifts from cut to cut.
Whooshes, impacts, risers and reverses, matched to your format palette, alongside an owned library for the hits that have to be exactly right.
This is the part nobody else does. A timeline engine places every stem to hit your exact length: the bed leads the voice in and ducks under it, SFX snap to word onsets from forced alignment, and the closing impact lands on the final syllable. Imaging effects — phone filter, stutter, reverse reverb, doubled voice — are applied as presets per format.
Each cut is mastered once at full resolution, then rendered to every delivery profile you selected. Loudness is applied as a single linear gain measured in a first pass, not a compressor riding the level, so the mix you approved is the mix that ships. Cuts too short for a reliable integrated reading are levelled on short-term loudness instead.
The finished take is transcribed back and compared to the script word by word, with call letters matched exactly. Length, clipping, dead air, loudness, voice-to-bed balance and a second copy review all run before anything reaches you.
Ten checks run before you hear a thing
A take that fails goes back to the stage that caused it and regenerates. If it runs out of retries it lands in our queue, not yours — and you are not charged until a passing take is delivered.
Transcript against the script: word error rate within tolerance, and the call letters and station name matching exactly.
The sung logo transcript contains the call letters or station name.
Final duration matches the target within tolerance.
No sample above the profile true-peak ceiling after mastering.
No internal gap longer than the limit, and no leading silence.
Integrated loudness within the delivery profile tolerance.
Voice sits clearly above the bed during speech, measured on the stems.
For legal IDs, the transcript contains the locked wording in order.
Spectral check for emergency-alert attention tones in the bed and SFX.
Second model pass flags awkward, offensive or competitor-naming copy.
Mastered for where it actually airs
One master at 48 kHz / 24-bit, then a render per delivery profile you select. Television follows the regulated loudness standards. Radio has no regulated target, so we match the music library you upload and sit imaging slightly hotter, the way a programmer would.
Full delivery specificationsWhy WAV, not MP3
Your air chain already encodes: the studio-to-transmitter link, HD Radio, DAB+, satellite. Each lossy stage stacks artifacts on the last. Masters stay uncompressed so the only encoder that touches the audio is yours.
What is in the file
BWF bext with station, product and origination date. AES46 cart
chunk with title, category, cut ID, start and end dates for dated promos, and
intro and end-of-message markers. iXML carrying the order and take ID, so a
file found loose in your automation can be traced back to the order.