Models
A model defines the type of processing applied to your audio. Reference a model by its alias in a job target — and combine several models in a single job. Input lists the channel layouts a model accepts, and a tier can accept fewer than the model does, so check the note under a model before sending 5.1; see Audio formats for the 5.1 requirements.
Dialogue, Music & Effects (DME)
Split a mixed soundtrack into dialogue, music, and effects for dubbing, localization, and re-mixing. Request one or more of these tracks per job. DME models also accept 5.1 (six-channel) input and return 5.1 output.
Processing profiles
DME targets support processing tiers per model. Check each model's supported tiers below; pass tier in your job to choose one. Separation results are content-dependent; compare Premium and Standard on representative samples when the exact result matters.
Applies an additional refinement pass where supported, which can improve results for some source material but takes longer to process. Results are source-dependent, so compare with Standard on representative samples when the choice matters.
Balanced processing profile that can be faster and more cost-conscious, and may better match the desired output for some source material or sections.
The causal, low-latency model built for live and streaming work. It is far smaller and quicker than the other profiles, and its separation quality is below Standard on dense or complex material. Today it is requested through the same file-based job API as every other tier — a streaming interface is not available yet. It accepts mono and stereo only, not 5.1, because it processes one canonical 48 kHz stereo signal.
dme_dialogue_v1— DialogueDefault voice extraction that keeps all voices, including dialogue and vocals, while removing background music and effects. Typically delivers cleaner voice extraction than the dialogue-only model.
dme_dialogue_remove_vocals_v1— Dialogue — remove vocalsDialogue-only extraction that removes vocals, music, and effects. Use it when sung or music vocals overlap dialogue and you need the dialogue isolated from those vocals.
live accepts mono and stereo only — 5.1 is not supported on this tier, even though the model itself accepts it.
dme_music_v1— MusicExtracts music, including vocals, while removing dialogue and effects.
dme_effects_v1— EffectsExtracts effects only, removing dialogue, music, and vocals.
dme_music_effects_v1— Music & EffectsPreserves music, effects, and vocals while removing only dialogue. Together with `dme_dialogue_remove_vocals_v1` it reconstructs the original mix.
dme_music_effects_remove_vocals_v2— Music & Effects — remove vocals v2Preserves music and effects only, removing both dialogue and vocals, with cleaner separation and higher fidelity than `dme_music_effects_remove_vocals_v1`.
dme_music_effects_remove_vocals_v1— Music & Effects — remove vocalsPreserves music and effects only, removing both dialogue and vocals. Typically delivers cleaner music-and-effects reconstruction than `dme_music_effects_v1`.
Stem Separation
Break a music track into individual music stems, including vocals and instruments. Each model produces one stem; request several model aliases in one job to receive multiple stems.
stem_vocal_v1— VocalExtracts the vocal stem from the multi-stem separation family. Use `karaoke_vocal_v1` instead when you want the karaoke-style vocal/accompaniment split.
stem_drum_v1— DrumExtracts the drum and percussion stem from music.
stem_bass_v1— BassExtracts the bass stem from music.
stem_electric_guitar_v1— Electric guitarExtracts the electric guitar stem from music.
stem_acoustic_guitar_v1— Acoustic guitarExtracts the acoustic guitar stem from music.
stem_acoustic_piano_v1— Acoustic pianoExtracts the acoustic piano stem from music.
Instrumental Track (Karaoke) Separation
Split a song into vocals and instrumental backing for karaoke, remixing, and vocal removal workflows. Each model produces one output stem.
karaoke_accompaniment_v1— AccompanimentExtracts the instrumental backing track from music.
karaoke_vocal_v1— VocalExtracts the vocal track from the karaoke-style vocal/accompaniment split while removing the instrumental backing.
De-noise
Separate clean speech from background noise for dialogue cleanup, podcasts, and voice recordings. Each model produces one output stem.
denoise_speech_v1— SpeechIsolates clean speech from background noise for dialogue cleanup, podcasts, and voice recordings.
denoise_noise_v1— NoiseExtracts the removed background noise from speech recordings.