babble¶

Generate babble speech from project files that use the babble label convention.

Overview¶

Babble generation now works on labeled talker groups instead of arbitrary file picks. Use labels like bab-f1, bab-f2, bab-m1, and bab-m2 to mark files that belong to a given talker group.

Both CLI and GUI call the same core function in triton.degrade.noise_generator so behavior stays consistent across interfaces.

The babble workflow is:

select the number of talker groups to mix
optionally set female and male counts separately
normalize each source file to the target RMS
concatenate all files for each selected talker group
honor an intended per-talker length so extra files are skipped when not needed
if a talker is short, repeat that talker's files randomly until target length is reached
mix the resulting talker tracks and optionally peak-normalize the output

When Babble output is added back into a project from the GUI, Triton stores it as a derivative artifact labeled bab-t#, where # is the number of talkers used. The derivative is written with a provenance sidecar that records the source files and babble generation parameters.

This is intended for speech-in-noise research, cocktail-party scenarios, and transcription stress testing.

Label Convention¶

Babble looks for file labels that match:

bab-f<number> for female talker groups
bab-m<number> for male talker groups

Examples:

bab-f1
bab-f2
bab-m1
bab-m2

If multiple files share the same label, they are treated as one talker and concatenated in filename order after per-file RMS normalization.

CLI¶

pixi run triton babble mix <project_dir> \
  --num-talkers <N> \
  [--num-female-talkers <F>] \
  [--num-male-talkers <M>] \
  [--intended-length-seconds 30] \
  [--target-rms 0.1] \
  [--peak-normalize / --no-peak-normalize] \
  [--output-path outputs/babble.wav]

Options¶

--num-talkers: Total number of talker groups to mix
--num-female-talkers: Number of female talker groups to use
--num-male-talkers: Number of male talker groups to use
--intended-length-seconds: Per-talker target duration; extra files are skipped once target is reached
--target-rms: RMS applied to each source file before concatenation
--peak-normalize: Normalize the final mix to safe headroom
--output-path: Output file path, default outputs/babble.wav

If neither female nor male counts are provided, Triton balances them as evenly as possible. For an odd total, one sex gets the extra talker group.

Example: balanced mix¶

pixi run triton files label my-project speaker_f1.wav "bab-f1"
pixi run triton files label my-project speaker_f2.wav "bab-f2"
pixi run triton files label my-project speaker_m1.wav "bab-m1"
pixi run triton files label my-project speaker_m2.wav "bab-m2"

pixi run triton babble mix my-project \
  --num-talkers 3 \
  --output-path outputs/babble_3talkers.wav

Example: explicit female/male split¶

pixi run triton babble mix my-project \
  --num-talkers 4 \
  --num-female-talkers 2 \
  --num-male-talkers 2 \
  --intended-length-seconds 45 \
  --target-rms 0.1 \
  --output-path outputs/babble_4talkers.wav

Workflow¶

Import the source files into a project, optionally assigning a single label to the full upload batch.
Label talker-group files with bab-f1, bab-m1, and similar labels.
Run triton babble mix with a total talker count.
Optionally set female and male counts explicitly.
Optionally set an intended per-talker duration.
Download the mixed babble output.

Python API¶

from pathlib import Path

from triton.degrade.noise_generator import generate_project_babble

result = generate_project_babble(
    Path("my-project"),
    sr=16000,
    channel_mode="mono",
    num_talkers=3,
    intended_length_seconds=30.0,
    target_rms=0.1,
    peak_normalize=True,
)

babble = result.audio
labels_used = [g.label for g in result.selected_groups]

Notes¶

Per-file RMS normalization happens before concatenation.
Multiple files can share a babble label; they are normalized, concatenated, and treated as a single talker group.
Intended-length planning avoids loading extra files when enough material already exists.
If a talker does not have enough unique source material, files are repeated randomly to reach target length.
Talker groups are selected from the project metadata, so files do not need to be copied anywhere else.
GUI and CLI use the same core babble generation function and therefore share identical selection and mixing behavior.
GUI babble generation can also persist the output back into the project with a sidecar so the derivative remains traceable.