Getting started with superassp

superassp is a self-contained R package for speech signal processing. It bundles a large collection of DSP routines (pitch, formants, voice quality, prosody, …) behind one consistent, wrassp-like interface, and produces output that plugs directly into emuR. No external Praat or wrassp installation is required.

The two function families

Almost everything you call falls into one of two prefixes:

Prefix	Produces	Output class	Example
`trk_*`	A time-series track following the audio (F0, formants, RMS, …)	`AsspDataObj`	`trk_pitch_rapt()`
`lst_*`	Summary statistics (scalars/vectors: jitter, voice-quality scores, …)	`JsonTrackObj`	`lst_voice_report()`

Two more families round things out: ucnv_* for unit conversions (Hz↔︎Bark/Mel/ERB/semitone, dB↔︎phon/sone) and read_*/write_* for I/O (read_audio, read_ssff, read_jstf, and their write_ counterparts).

Output modes: in memory vs. to file

Every trk_* and lst_* function takes a toFile argument:

toFile = FALSE — return the result in memory (an AsspDataObj or JsonTrackObj) for immediate use in R. Single file only.
toFile = TRUE (the default for most functions) — write an SSFF track file (trk_*) or JSTF file (lst_*) next to each input, and return a count. This is the batch mode; pass a vector of paths and it parallelises across files.

wav <- system.file("samples", "sustained", "a1.wav", package = "superassp")

# In-memory analysis
rms <- trk_rms(wav, toFile = FALSE, verbose = FALSE)
rms
#> In-memory Assp Data Object
#> Format: SSFF (binary)
#> 805 records at 199.547511312217 Hz
#> Duration: 4.034127 s
#> Number of tracks: 1 
#>   RMS[dB] (1 fields)

Inspecting an `AsspDataObj`

Use the accessor generics rather than digging into attributes:

track_names(rms)       # which tracks are present
#> [1] "RMS[dB]"
sample_rate(rms)       # frame rate in Hz
#> [1] 199.5475
n_records(rms)         # number of analysis frames
#> [1] 805
signal_duration(rms)   # seconds
#> [1] 4.034127

as.data.frame() (or as_tibble()) flattens all tracks into one time-indexed table — the usual bridge to dplyr and plotting:

df <- as.data.frame(rms)
head(df)
#>    frame_time   RMS_dB
#> 1 0.002505669 26.59310
#> 2 0.007517007 30.23827
#> 3 0.012528345 31.27690
#> 4 0.017539683 30.04947
#> 5 0.022551020 26.95653
#> 6 0.027562358 23.09678

Individual track matrices are reached by name, e.g. rms[["RMS[dB]"]].

Inspecting a `JsonTrackObj`

Summary (lst_*) functions return a JsonTrackObj: a self-describing container of measures with their schema and provenance.

vq <- lst_voice_report(wav, toFile = FALSE)
vq                       # compact summary
names(vq$field_schema)   # available measures
vq$slices[[1]]$values    # values for the first slice

Which pitch tracker should I use?

All pitch trackers emit a fundamental-frequency track (0 marks unvoiced frames). They differ in method, speed, and robustness:

Function	Method / origin	Notes
`trk_pitch_rapt`	RAPT autocorrelation (SPTK)	Fast, robust general-purpose default
`trk_pitch_swipe`	SWIPE′ (SPTK)	Accurate, good on noisy speech
`trk_pitch_reaper`	REAPER (Google)	Also yields voicing / GCI information
`trk_pitch_yin`	YIN	Classic difference-function tracker; +probability
`trk_pitch_pyin`	probabilistic YIN	Smoother voicing decisions than YIN
`trk_pitch_crepe`	CREPE deep neural net (ONNX)	Most accurate; heavier (downloads a model)
`trk_pitch_srh`	Summation of Residual Harmonics	Strong under additive noise
`trk_pitch_dio` / `trk_pitch_harvest`	DIO / Harvest (WORLD)	Fast (DIO) vs. accurate (Harvest)
`trk_pitch_ac` / `trk_pitch_cc` / `trk_pitch_shs` / `trk_pitch_spinet`	Praat (autocorrelation / cross-correlation / subharmonic / SPINET)	Praat-faithful behaviour
`trk_pitch_ksv` / `trk_pitch_mhs`	ASSP (K. Schäfer-Vincent / Michel)	Bundled C library, no external deps

Start with trk_pitch_rapt (fast, reliable). Reach for trk_pitch_swipe or trk_pitch_crepe when you need maximum accuracy, and trk_pitch_srh for noisy recordings.

Which formant tracker should I use?

Function	Method / origin	Notes
`trk_formant_burg`	Praat Burg LPC (via pladdrr)	Praat-faithful default
`trk_formant_forest`	ASSP forest tracker	Bundled C, no external deps
`trk_formant_snack`	Snack ESPS-style LPC	Familiar to Snack/WaveSurfer users
`trk_formant_cgdzp`	Complex-group-delay (COVAREP)	Robust bandwidth estimates
`trk_formant_tvwlp`	Time-varying weighted LP	Better on rapid transitions
`trk_formant_deepformants` / `trk_formant_formantnet`	Deep neural nets (ONNX)	Highest accuracy; heavier

Start with trk_formant_burg or trk_formant_forest. Use the neural trackers (trk_formant_deepformants, trk_formant_formantnet) when accuracy matters more than runtime.

Batch processing to files

files <- c("speaker1.wav", "speaker2.wav", "speaker3.wav")

# Writes one .rap SSFF file next to each input, parallelised across files
trk_pitch_rapt(files, toFile = TRUE)

The written SSFF/JSTF files load straight back with read_ssff() / read_jstf() and are ready for use with emuR.