Module synth 

The ftts say synthesis path: text in, 24 kHz PCM out.

This module is where the CLI stops describing the pipeline and runs it. It resolves a checkpoint bundle, hydrates the talker and the codec, tokenizes and wraps the text, derives the prompt header, drives TtsEngine::synthesize over the real QwenGenerator, and hands the generated codes to the codec decoder. What comes back is f32 samples; the WAV writing lives in ftts-core::audio and the sink policy in crate::AudioOutput.

Why the text is prepared before the engine runs

TtsEngine::synthesize owns text preparation, and normally that is where tokenization happens. Here it happens once, up front, and the engine is handed a preparer that returns that exact result. The reason is the cold text embedding: it is [151936, 2048], and materializing it whole to serve a fifteen-token utterance would cost 1.24 GB. The gather needs the token ids, the generator needs the gathered table, and the generator must exist before synthesize is called — so the ids have to be known first. The engine still receives, verbatim, the PreparedText a fresh call would have produced; nothing is skipped, only ordered.

Speaker conditioning is derived, never invented

An x-vector prompt conditions on a 1,024-wide speaker embedding. A voice source may be either a precomputed raw vector (1,024 little-endian f32, 4,096 bytes) or reference audio decoded through the pinned 24 kHz log-mel front end and ECAPA encoder. Neither path accepts a fabricated vector.

Structs

DenoiseReport

Derive an x-vector from an enrolled raw vector or a real reference recording. What a --denoise enrollment measured, so the CLI can report the effect rather than assert it.

DereverbReport

What a --dereverb enrollment measured, so the CLI reports the effect rather than asserting it.

LoadedModel

Every weight and table one say needs, hydrated once.

ModelBundle

The model resources ftts say needs, located relative to one model path.

ReferenceCleanup

Which reference-cleanup stages to run, and where each reports what it measured.

SynthesizedAudio

A completed synthesis: the codes the talker produced and the audio they decode to.

Constants

DENOISE_ARTIFACT_RELPATH

Single-channel speech enhancement: MMSE-LSA gains, decision-directed SNR, OM-LSA presence weighting, over a noise floor initialised offline and then tracked recursively.

SPEAKER_VECTOR_BYTES

Bytes in a speaker-vector file: 1,024 little-endian f32.

Functions

denoise_pcm_24k

Denoise arbitrary mono 24 kHz PCM through the pulled neural denoiser, if present.

generation_error

A model-side failure, for callers that need the engine’s own error type.

read_speaker_vector

Read a precomputed 1,024-wide speaker vector.

replace_speaker_vector

Replaces an existing enrolled voice, keeping the displaced one alongside it.

speaker_from_reference_pcm

Derive a speaker vector from already-decoded mono 24 kHz reference PCM.

speaker_from_voice

Derive a speaker vector from a voice source: a raw x-vector file, or reference audio.

synthesize

Run one utterance end to end: text, codes, PCM.

write_speaker_vector_new

Write a raw x-vector without replacing an existing enrollment result.