Effects

One kudio effect at a time, on the selection when there is one and on the whole clip when there is not. Choose the effect from Effect, set its parameters, and the line above the buttons shows the call Apply will make. That same call goes into the undo history, and from there into an exported script.

Preview, Apply, Undo

♪ Preview plays the effect over the selection, or the whole clip, and changes nothing: no edit, no undo step. It loops, so finding the right top_db is a matter of listening rather than apply, listen, undo, again. Change a parameter while it plays and it renders again, 120 ms after the last change. Preview and Apply run the same code on the same audio, so what you hear is what Apply will commit.

Anything else that claims playback ends the audition and releases the button: Play, Stop, an edit, another panel's preview.

Apply makes the edit. Undo takes back the last one and Revert to original drops them all; both are the same as in the Edit menu.

An effect applied to a selection replaces the range with a processed copy of itself, with no crossfade at its edges: blending them back would leave a few milliseconds at each end without the effect.

The effects

Parameter names are kudio's, so the call in the caption reads the same as the controls.

Level and fades

Effect	Parameters	What it does
gain	db	a fixed gain
random_gain	db_range low and high, seed	a gain drawn from the range
fade_in, fade_out	duration, shape	a fade at the start or the end; shape is linear, cosine or exponential
fade	fade_in_s, fade_out_s, shape	both at once
reverse		the selection backwards
remove_dc		subtracts the mean, per channel. An offset that drifts over a long recording is a very low frequency, which is a job for highpass

Filters and pitch

Effect	Parameters	What it does
highpass, lowpass	cutoff, order, zero-phase	a Butterworth filter. Zero-phase, on by default, runs it forwards and backwards so nothing moves in time
bandpass, bandstop	low, high, order	a band kept or taken out. bandstop starts at 45–55 Hz, around 50 Hz mains hum
pitch_shift	n_steps	up or down, in semitones, keeping the length
time_stretch	rate	faster or slower, keeping the pitch; the length changes

A cutoff cannot go above half the clip's sample rate. The controls stop there, rather than letting Apply fail on it.

Rooms

Effect	Parameters	What it does
reverb	decay, delay_ms	an echo every delay_ms, each decay times the one before: cheap, and not a room
room	rt60, drr_db, seed	a generated impulse response, from kudio.rir: how long the tail takes to fall 60 dB, and how loud the direct sound is against it
measured_room	an impulse response file, start at the direct sound	a room someone measured; below

Noise and channels

Effect	Parameters	What it does
add_noise_snr	a noise file, snr_db, seed	noise from a file, mixed in at the SNR asked for
telephone	codec	a phone line: band-limited and resampled, with G.711 μ-law or A-law companding, or none to hear the line without the codec
bit_depth	bits, dither, seed	fewer bits. With dither ticked, a triangular (TPDF) dither two steps wide, peak to peak, goes in before rounding
dropouts	loss, packet_ms, burst, conceal, seed	lost packets, as on a VoIP link; below

Silence and speech

Effect	Parameters	What it does
trim_silence	top_db, frame_length, hop_length	cuts quiet audio off both ends, by level
vad_trim	threshold_db, pad	cuts both ends down to the first and last speech kudio.vad finds; below
split_on_silence	top_db, frame_length, hop_length	finds the non-silent stretches without changing anything; below

Seeds

A seed of −1 shows as None or random, and draws a new result every time. Any other value gives the same result every time, so a room or a dropout pattern can be rebuilt later. Apply on a random seed still records the seed it drew, and an exported script repeats that result rather than rolling again.

A room someone measured

room builds a statistical room from an RT60 and a DRR: reproducible, a few controls, and no early reflections. That is what augmentation wants, and not what a particular hall sounds like. measured_room takes an impulse response from a file instead and applies it with kudio.apply_rir. The file is read the way the recorded call reads it, with kudio.file_load at the clip's rate and in mono, and read again if the clip changes rate, so Preview, Apply and an exported script all use the same samples:

y = kudio.apply_rir(y, kudio.file_load('/rooms/hall.wav', sr=sr)[0][192:])

The [192:] is start at the direct sound, on by default. A measured response begins with the time the sound took to reach the microphone, and left in, that delays the audio: a processed selection comes back late against its neighbours. This starts the response at its loudest sample, which is the direct sound in any ordinary room, and where kudio.rir puts it.

Under the file, kudio.rt60 reads the response as it will be applied, resampled, mixed to mono and trimmed, so a hall measured at 48 kHz and applied at 8 kHz is described as what is left of it:

0.91 s at 16000 Hz, direct sound at 12.0 ms — applied from there
T30 0.59 s (EDT 0.54 s)  ·  C50 +9.4 dB  ·  DRR +3.0 dB  ·  fit 1.000

A file that is not an impulse response still produces a T30. When the decay is not a straight line, or its EDT and T30 disagree, the reading says so and turns red.

Lost packets

dropouts drops whole packets, the way a VoIP link loses them.

• loss is the fraction of packets dropped, and it is achieved exactly, so a corpus labelled 5% carries 5%.
• packet_ms is the packet length, 20 ms by default as in G.711 over RTP. It decides whether the damage comes as many short holes or a few long ones, not how much of it there is.
• burst is the mean run length of a loss. Real links lose packets in runs, and a run is far harder to conceal than the same packets scattered, because there is no recent speech left to repeat. independent is not runs of one: independent draws land side by side sometimes, and already average 1/(1 − loss), 1.05 packets at 5%.
• conceal fills a hole with silence, or with hold, which repeats the last packet that arrived. Whether hold measures better than a hole depends on whether the repeat lands in phase, an accident of the talker's pitch against the packet length, so a waveform metric cannot choose between them. Across a burst it tiles one packet over the whole gap, and there it is plainly worse.

Speech, not level

trim_silence and split_on_silence cut on level alone. That works on a studio take and falls apart on anything with a fan in the background: above some noise level no fixed top_db both keeps the quiet consonants and drops the room.

vad_trim decides frame by frame against the recording's own noise floor, with kudio.vad, and uses spectral flatness to tell a voice from a hiss. pad keeps a little audio either side of the speech, and a clip with no speech found in it comes back untouched rather than empty. The same detector is behind Find speech in the marker list.

Splitting on silence

split_on_silence cannot be previewed, because it changes nothing. Apply reports how many non-silent segments it found and how long they are in all, and Save segments… then writes each one to a folder as its own WAV file, numbered after the clip: clip_001.wav, clip_002.wav and so on.