Module engine 

Core TTS engine trait.

Traits

TtsEngine

Trait that every TTS engine must implement.

Functions

estimate_word_boundaries

Estimate word boundaries using word-length-adjusted timing with an inter-sentence pause. Originally mirroring Swift’s WordTimingEstimator, since extended (sentence gaps) beyond it.

estimate_word_boundaries_with_wpm

Estimate word boundaries with configurable words per minute. Abbreviation caveat: token-level detection means “Mr.” / “e.g.” also get the sentence gap — an accepted trade-off for an estimator.

preprocess_speech_markdown

Convert SpeechMarkdown to SSML when detected, otherwise return the text unchanged. Returns (processed_text, is_ssml).

strip_ssml_to_text

Strip all XML/SSML tags from ssml, returning the plain-text content with collapsed whitespace. Used to convert incoming W3C SSML (from tts_speak_ssml) to plain text for engines that don’t accept SSML (OpenAI, ElevenLabs, SherpaOnnx, etc.).

syllable_count

Rough syllable count for duration weighting: contiguous vowel groups (English-centric), every digit spoken individually, minimum one.

trailing_pause_weight

Pause weight AFTER a word, in syllable-equivalents, from its trailing punctuation. Sentence ends breathe longest; commas/semicolons beat.

unwrap_voice_tag

Extract the name attribute from a <voice name='...'> (or "...") tag and return the inner SSML content wrapped in a bare <speak> element.

weighted_boundaries_scaled

Word boundaries with RELATIVE durations mapped onto a known total duration (engines that know the audio length but not per-word times — e.g. qwen3-local). The word shape comes from the syllable/pause weights; the timeline is linearly rescaled so the last word ends exactly at total_secs. Offsets/durations are absolute seconds.

word_weight

Duration weight of a word in syllable-equivalents: syllables plus a small constant for onsets/codas so “a” is not zero-length.

Type Aliases

OnAudioCallback

Callback for streaming audio chunks.

OnBoundaryCallback

Callback for word boundary events. Signature: (word, start_sec, end_sec, byte_offset, byte_len, estimated) byte_offset is a byte index into the text as submitted (for SpeechMarkdown input, the caller’s original source — highlighting targets the caller’s string) and byte_len a byte length. When a word cannot be located (normalization mismatch, punctuation artifact), the callback receives the last known offset held (0 if nothing matched yet) with byte_len -1 — not a -1 offset. estimated is true for proportional estimates (unpatched voices, wpm model), false for measured timings (cloud provider timings).

OnEndCallback

Callback for speech-finished events.

OnErrorCallback

Callback for error events.

OnMarkCallback

Callback for SSML mark/bookmark events. Signature: (name, start_sec, end_sec, char_offset) char_offset is -1 when the engine doesn’t report it. Engines that estimated) audio position; consumers map it to their bookmark event (SAPI SPEI_TTS_BOOKMARK, SSIP index marks).

OnStartCallback

Callback for speech-started events.