Naming Conventions

About 4 min

Naming Conventions

Naming matters a lot in ESPnet3.

Good names make configs easier to read, stage flows easier to follow, and code search much more reliable for both humans and coding agents.

This page summarizes the naming patterns currently used in the ESPnet3 codebase and docs.

Important

Prefer existing names over inventing new ones. If the repository already has a stable term for a concept, reuse that term across code, configs, docs, and file paths.

General rule

A good ESPnet3 name should be:

consistent with existing code
easy to search
specific about its role
stable across code, config, and docs

If a new name would create a synonym for an existing concept, do not add it.

System directory names

Apply these rules, in order, to names under espnet3/systems/:

Use lowercase.
Keep the name clear and unambiguous.
Use the shortest name that remains clear. For a system directory, established domain abbreviations are appropriate when they prevent an unwieldy name: esp2, asr, tts, st, and enh are accepted terms. This is a deliberate exception to the general preference for full words in variables, files, and APIs.
Prefer one word where possible. Use underscores only when they preserve an established name, version, variant, or meaningful structure.

ESPnet2 task-derived systems use esp2_<task>, such as esp2_asr and esp2_asr_transducer. The short prefix keeps directory names manageable while grouping their ESPnet2 provenance in alphabetical listings. ESPnet2-derived work without a task-shaped system remains a regular system name, such as speechlm.

Examples of system directory names

Name	Why it fits
`f5tts`	Established one-word system name.
`openbeats`	Established one-word system name.
`parakeet`	Clear one-word system name.
`granitespeech`	Clear one-word system name.
`speechlm`	Established name; no separator is needed.
`whisper`	Established one-word system name.
`wav2vec2`	Established one-word system name.
`wavlm`	Established one-word system name.
`hubert`	Established one-word system name.
`seamlessm4t`	Established one-word system name.
`owsm_v3`	The underscore preserves a meaningful version.
`owsm_v4`	The underscore preserves a meaningful version.
`qwen2_audio`	The underscore preserves the established model family structure.
`qwen2_5_omni`	The underscores preserve the established version and variant.
`esp2_asr`	ESPnet2 ASR task-derived system.
`esp2_asr_transducer`	ESPnet2 ASR transducer system; the short `esp2` prefix keeps it manageable.
`esp2_tts`	ESPnet2 TTS task-derived system.
`esp2_st`	ESPnet2 speech translation task-derived system.
`esp2_enh`	ESPnet2 speech enhancement task-derived system.
`huggingface_asr`	A toolkit plus task clearly identifies a wrapper system.

Python naming

Follow normal Python naming first:

functions and variables: lower_with_under
classes: CapWords
constants: CAPS_WITH_UNDER
module files: lower_with_under.py

For internal helpers, use a leading underscore:

_normalize_stage_name
_load_inference_config

Function names should read like actions

In practice, ESPnet3 function names usually start with a verb or an action-oriented prefix.

Good examples:

load_config_with_defaults
resolve_stages
run_stages
collect_stats
prepare_sentences
upload_model

Avoid vague noun-like function names such as:

stage_config
dataset_module
model_state

If the function does something, the name should sound like that action.

Note

Common ESPnet3 action prefixes include load, build, run, collect, validate, prepare, infer, measure, pack, and upload.

Class names should read like entities

Class names should look like nouns, roles, or objects.

Good examples:

BaseSystem
InferenceRunner
DatasetBuilder
ESPnetLightningModule

Avoid class names that read like commands:

LoadDataset
RunInference
BuildTrainer

If the code represents a thing, the name should read like a thing.

File names should also read like entities

Python file names should also look like nouns or role names, not verb phrases.

Good examples:

system.py
inference_runner.py
dataset_builder.py
config_utils.py

Avoid names like:

run_stage.py
load_config.py
prepare_dataset.py

When a file holds a role or domain concept, name it after that concept.

Prefer inference, not decode

For new ESPnet3 code, do not introduce decode when the concept is inference output or inference-time behavior.

Historically, decode was a natural name for ASR-style pipelines. But ESPnet3 now covers more generative systems, and many of those workflows are not best described as "decoding". inference is the broader and more stable term, so new naming should prefer that direction.

Prefer names such as:

inference
inference_dir
inference_config
InferenceProvider

Avoid new names such as:

decode_dir
decode_config
decode_output

This rule applies to:

config keys
path names
stage-facing field names
publication settings
docs and examples

Warning

Older code may still contain decode in some places. Do not copy that naming into new ESPnet3 work unless you are intentionally working inside a legacy compatibility boundary.

Use snake_case for stage names and config keys

Stage names should be short verb-style snake_case names:

create_dataset
train_tokenizer
collect_stats
prepare_labels
pack_model

Config keys and path-related fields should also use snake_case:

training_config
inference_dir
output_keys
stage_log_mapping

Do not mix styles such as camelCase or kebab-case into Python-facing config.

Imports under egs3/

Do not use relative imports inside egs3/.

Use explicit absolute module paths:

from egs3.mini_an4.asr.dataset.builder import MiniAn4Builder

Avoid:

from .builder import MiniAn4Builder

Absolute imports are easier to search, easier to move, and easier for coding agents to reason about.

Name by boundary

Use names that match the boundary where the code lives.

In shared espnet3/ code:

prefer reusable, system-level names
avoid recipe-specific corpus names

In recipe-local egs3/<recipe>/<system>/ code:

recipe-specific names are fine
names should still stay consistent with the shared API

Examples:

shared: InferenceRunner, DataOrganizer, BaseMetric
recipe-local: MiniAn4Builder, LibriSpeechDataset

Do not create unnecessary synonyms

Naming drift usually starts when two words mean the same thing.

Examples of bad drift:

infer vs decode
metric vs measure used inconsistently for the same concept
dataset_dir vs data_root vs corpus_dir for the same path

Before adding a new name, search the repository and reuse the dominant term if the meaning is the same.

Avoid unnecessary abbreviations

Prefer full words unless the abbreviated form is already widely understood in the field or already well established in ESPnet3.

Good:

collect_stats
inference_dir
config_utils

Usually avoid:

prep_data
cfg_loader
infer_out

Prefer:

prepare_data
config_loader
inference_output

Short forms are fine when they are already common and unambiguous. stats is a good example. Ad-hoc shortening is not.

Note

If a shortened name saves only a few characters but makes the meaning less obvious, use the full word.

A practical naming workflow

Before introducing a new name:

Search for the concept in espnet3/ and egs3/.
Reuse an existing name if one already matches.
Check whether the name matches the expected Python style.
Check whether the name is specific to shared code or recipe-local code.
Verify the actual file or directory name before writing links or imports.

Examples

Good:

def load_inference_config(path: Path) -> DictConfig:
    ...


class InferenceProvider:
    ...

Good:

stage_log_mapping = {
    "infer": "inference_config.inference_dir",
    "measure": "metrics_config.inference_dir",
}

Bad:

def config_loader(path: Path) -> DictConfig:
    ...


class RunInference:
    ...

Bad:

decode_dir = "exp/my_run/decode"
from .builder import MyBuilder

Adding a stage

See how stage names and custom system methods fit together.

Docstring guide

Write documentation that matches these naming choices.

CI and PR

See review expectations for naming and cleanup changes.