One conversation, two interventions

What enters and leaves

The research pipeline accepts long-form speech. It produces a transformed conversation, transcript/replacement artifacts, speaker assignments, and stage-specific measurements. The framework seeks to change vocal identity and protected lexical content while retaining useful surrounding evidence. Preservation must be measured, not assumed.

Identity transformation

1. Diarization and transcription estimate who spoke when and what was said. Errors here propagate into speaker profiles, target assignment, lexical planning, and editing.
2. Speaker profiling and retrieval build embeddings and use available metadata to choose voice references. Source-to-reference distance and reference-to-reference distinctiveness are separate constraints. A consistent assignment is reused for each source speaker throughout a conversation. A retrieval distance is not an ASV EER.
3. SeedVC converts every speaker turn to its assigned voice. It does not detect identifiers or decide which words should change.

Lexical transformation

1. ASR and LLM planning produce a transcript and document-consistent synthetic replacements in the transcript's language. Detection and planning are distinct from whether the planned words are actually spoken later.
2. Complete-turn rendering gives the editor the converted speaker turn, its complete source transcript, and its complete revised transcript. F5-TTS synthesizes a turn; VoiceCraft uses audio/text conditioning to edit speech. Only affected turns are lexically rerendered. The framework does not define F5-TTS as the entire system.
3. Reconstruction inserts rendered turns into converted context. The non-edited context must come from the voice-converted audio, not the original identifiable voice.

Public interfaces

The package exposes two levels:

• Service components: profiles, retrieval, conversion, ASR, planning, editing, artifact storage, events, queue, and control-plane API.
• Complete-turn assembly: a narrow renderer-independent interface for a converted waveform plus revised turn transcripts. It preserves intervening samples and records exact output intervals and hashes. This interface does not run the upstream stages.

The service engine retains configurable historical research modes. Its default workflow is not the exact fixed paper experiment. Use explicit backend, model, language, and policy settings, and label baseline/ablation configurations. Do not silently substitute legacy runtime quality proxies for the clinical study endpoints.

Evaluation is a separate path

Score original audio, voice-converted audio, generated edits, and final reconstructed speech separately. Keep failures and uncertainty in the registered population. The complete-turn interface never uses an ASR score to reject an otherwise valid WAV or silently replace it with source speech. Structural failures stop assembly visibly.