Voiceover production and AI voice generation each have their own vocabulary, and working with AI voices means using
both. Here are plain definitions of the terms that come up most, from ADR to zero-shot, with links to
the guides where they matter.
A to D
ADR (automated dialogue replacement)
Re-recording dialogue after filming, in a studio, to replace or fix lines recorded on set.
Bark
A short line a game character says in reaction to something: spotting an enemy, reloading, greeting the player. See writing game barks.
CFG (classifier-free guidance)
A technique that pushes a model's output more strongly towards its inputs. In abcVoiceover it's guidance strength: how closely a take follows the voice clip, especially its pacing.
Clipping
Distortion when a recording is louder than the maximum level the system can store. It sounds harsh and crackly, and can't be repaired.
Condenser microphone
A sensitive microphone that captures detail, and the room too. See choosing a microphone.
dBFS
Decibels relative to full scale: digital level, where 0 is the loudest possible and everything else is negative.
Dynamic microphone
A less sensitive microphone that picks up mostly what's close to it, useful in untreated rooms.
E to L
Emotion intensity
abcVoiceover's control over how dramatic a voice's delivery is, from flat to theatrical. Chatterbox calls it exaggeration.
Expressiveness
abcVoiceover's name for sampling temperature: how varied each take is. Low is steady and predictable; high is livelier and less predictable.
Fine-tuning
Training an existing model further on specific data, such as hours of one person's voice. The alternative to zero-shot cloning.
Gain
How much a microphone's signal is amplified on the way in. Too high causes hiss and clipping; too low makes a weak recording.
IVR (interactive voice response)
An automated phone menu: "For sales, press one." See phone menus.
Line
In abcVoiceover, one unit of a script: a piece of text, the voice that reads it, and its take. Each line becomes one audio file.
LUFS
Loudness units relative to full scale: a measure of how loud audio sounds over its length, used for platform loudness targets. See loudness standards.
M to P
Min P (min probability)
A sampling limit that rules out unlikely choices. Higher is safer and more predictable.
Noise floor
The level of background noise in a recording when nobody is speaking.
Normalization
Adjusting audio's level to a target, by peak or, better for voice, by loudness.
Pickup
A line recorded after the main session, to fix or add something.
Plosive
A burst of air on sounds like p and b that thumps a microphone. Avoid it by speaking slightly past the mic, or with a pop filter.
Prosody
The music of speech: rhythm, stress, intonation and pauses. Much of what makes a voice sound natural or robotic.
R to S
Reference clip (voice sample)
The recording a cloned voice is made from. In abcVoiceover, 5 to 15 seconds is recommended. See recording a voice sample.
Repetition penalty
A sampling setting that discourages repeating recent sounds. Keep it near 1.2; set high, it makes the voice ramble past the end of a line.
Room tone
The sound of a room with nobody speaking, recorded to fill gaps naturally. Some audiobook platforms require a little at the start and end of each file.
Sample rate
How many times per second audio is measured, in kHz. abcVoiceover's takes are 24 kHz; music is usually 44.1 or 48 kHz; phone systems often use 8 kHz.
Scratch track
A temporary recording used while editing or animating, replaced later by the final one. See scratch dialogue.
Sibilance
Harsh "s" and "sh" sounds, reduced with a de-esser or better mic placement.
Speech tokens
Small numbered units that represent slices of sound, which a voice model generates before turning them into audio. See how AI voice generators work.
T to Z
Take
One recorded or generated version of a line. In abcVoiceover you compare a new take with the current one and keep the better.
A sampling limit that only considers the most likely choices adding up to a given share. Lower is more conservative.
True peak
The highest point of a waveform, including between samples. Kept below about -1 dBTP to avoid distortion after conversion to MP3 or AAC.
Voice cloning
Producing new speech in a particular person's voice from a recording of them. See what is AI voice cloning?
Watermark (audio)
An inaudible signal hidden in audio to show where it came from. Chatterbox adds one to everything it generates.
WAV
An uncompressed audio file format. abcVoiceover takes voice samples and exports takes as WAV.
Whisper
An open-source speech recognition model. abcVoiceover uses it to check that every take says its words.
Zero-shot
Doing a task for something the model never saw in training. Zero-shot voice cloning copies a voice from one short clip with no training. See zero-shot voice cloning.
Can we use Google Analytics cookies to count visits to our public pages? It helps us see which
pages are useful. Nothing inside your projects is ever tracked.
Privacy policy