Extract Voiced Segments with AVQI Filtering (Tier 4 Ultra) — extract_voiced_segments

Complete AVQI voiced segment extraction pipeline in a single C++ call. Supports both AVQI v2.03 (simple intensity-based) and v3.01 (with ZCR filtering). Performs: Sound -> TextGrid (silence) -> Extract sounding -> Concatenate -> [v3.01: Window power/ZCR filtering] -> Final concatenation, in a single C++ call instead of a multi-step R implementation.

Usage

extract_voiced_segments_ultra(
  sound,
  version = "v3.01",
  min_pitch = 50,
  silence_threshold_db = -25,
  min_silent_duration = 0.1,
  min_sounding_duration = 0.1,
  power_threshold_factor = 0.3,
  max_zcr = 3000,
  window_width = 0.03
)

Arguments

sound: Sound object or external pointer
version: AVQI version: "v2.03" (simple) or "v3.01" (ZCR filtering, default)
min_pitch: Minimum pitch for silence detection in Hz (default 50)
silence_threshold_db: Silence threshold in dB (default -25)
min_silent_duration: Minimum silent interval duration in seconds (default 0.1)
min_sounding_duration: Minimum sounding interval duration in seconds (default 0.1)
power_threshold_factor: Power threshold as fraction of global power (default 0.3)
max_zcr: Maximum zero-crossing rate for voiced segments (default 3000)
window_width: Window width for v3.01 filtering in seconds (default 0.03)

Value

Sound object containing only voiced segments (concatenated)

Details

**TIER 4 ULTRA API**

This function implements the exact AVQI voiced extraction algorithm in a single C++ call. It replaces the R loops and multiple boundary crossings of the standard multi-step approach.

**Algorithm (AVQI v3.01):** 1. Detect silences using intensity-based TextGrid creation 2. Extract all "sounding" intervals 3. Concatenate sounding intervals -> "loud_sound" 4. Slide 30ms windows through loud_sound 5. Filter windows by: power > 30 6. Concatenate passing windows -> final voiced sound

**Algorithm (AVQI v2.03):** Steps 1-3 only (no window filtering)

**Version Differences:** - v2.03: Simpler, keeps most voiced content (~37s from 37s input) - v3.01: Aggressive ZCR filtering, removes fricatives (~25-30s from 37s input)

Algorithm choice

No pitch algorithm is used here — voiced/silence segmentation is Intensity-threshold based (`Sound_to_Intensity()` + `silence_threshold_db` relative to the maximum), not derived from a `Pitch` object. `min_pitch` only controls the Intensity analysis window length (via `Sound_to_Intensity`'s own floor parameter), not a pitch-detection algorithm choice. See the Tier 4 Ultra algorithm table in `inst/agents/AGENT_GUIDE.md`.

References

- AVQI v3.01: Maryn et al. (2017) - with ZCR filtering - AVQI v2.03: Original Praat script - intensity-based only

Examples

sound <- Sound$create_tone(frequency = 150, duration = 1, sampling_rate =
 16000)

# AVQI v3.01 (default, with ZCR filtering)
voiced_v3 <- extract_voiced_segments_ultra(sound, version = "v3.01")

# AVQI v2.03 (simple intensity-based)
voiced_v2 <- extract_voiced_segments_ultra(sound, version = "v2.03")

voiced_v3$get_total_duration()
#> [1] 0.001
voiced_v2$get_total_duration()
#> [1] 0.001