AudioGen Research

A 16-line audio drama scene: Viktor, a tired NY detective, calls Laura from the streets of Manhattan. Two voices, a telephone filter on her side, rain and city atmosphere, slow jazz vinyl bleeding through her apartment, and a final outro of his footsteps walking off.

The scene served as a controlled end-to-end test: generate assets, select takes, process dialogue, and arrange every element inside AudioGen. It tested a complete workflow against a concrete result rather than only demonstrating a UI.

Qwen3-TTS turns a prose description ("A weathered American male voice in his early forties, deep baritone with a slight gravel from too many late nights…") plus a 25–30 second sample text into a custom voice that gets saved into the Voice Library with language, gender and bucket. The library on the right shows speakers organized by project — TheBlackVeil, Charlie_Project, NOIR CALL — and you assign saved voices straight to the Script tab's Speaker 1 / Speaker 2 slots.

Direction controls such as [soft], [firm], [urgent], or [whisper] were tested per line. Their effect depends on the selected voice engine. The robust part of the experiment was the approval workflow: accept or regenerate individual takes without touching the rest of the session.

Neural TTS can miss: a clipped word, a dropped syllable, a sudden pitch swing, or a voice drifting from its reference timbre. AudioGen's Voice-Quality-Scoring combined several technical measurements into a 0–100 score and flagged unusual takes. The score was a search signal for human review, not objective proof of performance or audio quality.

Every line in the per-line approval queue carries its VCS number. Re-generations are compared against the previous best, and the spectrogram + outlier views make the failure mode visible instead of guessed.

Phone pickup, taxi pass on wet asphalt, footsteps: short, concrete one-shots were tested with MMAudio. Rain on rooftops, distant traffic, and room tone were created as longer loops with AudioLDM2. In testing, prompts focused on one clear element were more reliable than overloaded descriptions. Generated clips save into the SFX Library with category tags such as Ambience, Foley, Impact, or UI and can be placed directly on Composer tracks.

Six modes: Generate a fresh track, Extend an existing one, Add Intro / Add Outro, Repaint a drag-selected region without touching the rest, or Cover an existing track in a different genre. Style presets ("Lo-Fi Cafe", "Cinematic Build", etc.) prefill prompt + parameters with proven defaults that you edit on top. The lyrics field is optional — leave blank for instrumental, add [verse] / [chorus] / [bridge] tags for sung output.

Pick a session from the dropdown, see every line organized by speaker on the left, build an effect chain on the right. Apply a chain to an entire speaker (Laura gets the Telephone-Filter — High-Pass 300 Hz, Low-Pass 3400 Hz, slight saturation, compressor, gain trim — for the whole call) or override one specific line for emphasis. The Compare button A/Bs original vs. processed; Export FX Combined writes the rendered version back as a session-bound artifact.

Mastering chains use a separate preset category. “Music Master Streaming (-14 LUFS · Spotify/Apple)” was the tested starting point, combining high-pass filtering, presence processing, compression, limiting, and LUFS normalization. Below the music library, Separate Stems uses Demucs to split tracks into drums, bass, vocals, and other. The stems become children of the master track and can enter the Composer as four parallel lanes with independent volume, FX, and mute control, or be remixed into a new master.

Voice sessions auto-sync from the Script tab as two parallel speaker tracks with proper gaps between lines. SFX clips drag from the library onto their own lanes. Music sits on a dedicated lane (or four lanes if you dropped stems). Per-track color, type icon, drag-handle rename, individual volume. Auto-snap to clip edges, BPM beats and markers. Auto-crossfade on overlapping music / SFX / ambience clips. Multi-select with Alt+Drag, ripple-edit with Alt+Shift+Drag. Two export targets: a single WAV / MP3 of the audio mix, or an MP4 with the audio plus audio-reactive visuals (spectrum bars, waveform pulse) overlaid on a still cover image.