> ## Documentation Index > Fetch the complete documentation index at: https://api-docs.ollang.com/llms.txt > Use this file to discover all available pages before exploring further. # Audio Processing Providers > Stem separation in the Ollang dubbing pipeline — how the dialogue stem is isolated from music and effects, and what that means for the final mix. Audio processing runs in the `split` step, the first stage of an **AI Dubbing** Order. It separates the source audio into an isolated dialogue stem and a music-and-effects (M\&E) bed. That separation is what makes dubbing sound like dubbing rather than a voiceover. The dialogue stem feeds transcription; the M\&E bed is preserved and remixed under the synthesized target-language voice, so the score, ambience, and sound design survive localization intact. ```mermaid theme={null} flowchart LR A["Source audio"] --> B{"Stem separation"} B --> C["Dialogue stem"] --> D["Transcribe and translate"] D --> E["Synthesized voice"] B --> F["Music and effects bed"] --> G["Final mix"] E --> G ``` ## At a glance | Provider | Provider tag | Order types | Status | | -------------------- | ------------ | ----------- | ----------- | | [LALAL.AI](#lalalai) | `lalalAi` | AI Dubbing | **Default** | It is the shortest step in the dubbing pipeline, and the one with the largest effect on whether the finished mix sounds professional. *** ## LALAL.AI A dedicated stem separation service built on a set of proprietary neural networks — **Phoenix**, **Orion**, **Perseus**, **Andromeda**, **Lyra**, and **Lynx** — capable of splitting a track into twelve distinct stem types. An `auto` option selects the most effective network for the requested stem. The **Orion** network takes a different approach from conventional separation. Rather than carving stems out of the mix with spectral masks, it **synthesizes** them, reconstructing the stem from what it has learned across many mixes. The practical result is markedly less of the artifact residue that betrays a dub: the watery, phase-smeared remainder where dialogue used to be. **Best for** — every AI Dubbing Order. Dialogue-over-music content benefits most, since that is where mask-based separation degrades hardest. **Strengths** * Direct-synthesis separation with lower cross-stem leakage than mask-based methods. * **Three noise-cancelling levels** on the voice stem, for source audio of varying quality. * **Dereverberation**, which removes room echo from the extracted voice. This matters beyond the mix: reverb in the source is one of the more reliable ways to degrade transcription accuracy, so cleaning it here improves every downstream step. * **Separate `voice` and `music` stems** — spoken voice is a distinct target from sung `vocals`, which is exactly the split a dubbing pipeline needs. * **Selectable extraction level** (`deep_extraction` or `clear_cut`), trading completeness of the extracted stem against cleanliness of what is left behind. * Finer stems — drums, bass, piano, guitar variants, synthesizer, strings, wind — for mixes that have to be partially rebuilt rather than simply split. **Considerations** * Separation quality tracks source quality. Heavily compressed audio, dialogue buried under a loud mix, or material already processed by another separation tool will produce a weaker M\&E bed. * Where the source has a clean dialogue stem available, supplying it directly will always beat separating it back out. **In Ollang** — `lalalAi` · AI Dubbing · Default. *** ## Related The voice engines that render into the preserved M\&E bed. Supported source formats and how audio is ingested.