Speech Language Model

This template contains two drivers for different SpeechLM training interfaces:

Driver	Python entry point	Use
speechlm.sh	espnet2.bin.speechlm_train	Existing staged recipes, such as LibriTTS and mini_an4, with the traditional ESPnet environment and job launcher
train.sh	espnet2.speechlm.bin.train	Bagpiper and Bagpiper-TTS training with TorchTitan, prepared dataset manifests, and length statistics
scripts/speechlm/stage_utils.sh	sourced by those recipes' run.sh	Finding the latest complete checkpoint, and the rule for where a training stage starts
scripts/speechlm/parse_options.sh	sourced by train.sh and those recipes' run.sh	Reading the options, with repeatable ones collected into an array and every value kept literal

The drivers have different configuration and data interfaces. Keep using speechlm.sh for recipes built around it; train.sh provides the entry point for the current SpeechLM trainer. The existing setup.sh scaffolds the traditional speechlm.sh layout.

Recipes without data preparation

Bagpiper and Bagpiper-TTS contain run.sh, conf/, and a README. Each run.sh enters its recipe directory and writes its stages out in the usual egs2 shape, ending in export and inference:

	Bagpiper	Bagpiper-TTS
1	warmup	SFT
2	pretraining	export
3	SFT	inference
4	export
5	inference

./run.sh runs them in order and --stage / --stop-stage select part of that, so the last stage decodes published weights without training anything when it is given --export-path and --train-config. A training stage continues its own output directory once that has checkpoints, and otherwise starts from the stage before it; scripts/speechlm/stage_utils.sh holds that rule and the search for the latest complete checkpoint, so the recipes do not repeat it.

Their inputs are already prepared, so the data-preparation and cluster-launch files are unnecessary:

• Environment: there is no path.sh to activate tools/venv. Activate the environment from the SpeechLM installation guide first, or pass --python /path/to/env/bin/python. train.sh adds the repository root to PYTHONPATH and uses that Python interpreter.
• Launch: there is no cmd.sh, ${cuda_cmd}, run.pl, or slurm.pl dispatch. train.sh invokes torchrun directly. For multiple nodes, start the command on every allocated node with the same --num-nodes, --master-addr, and --master-port, and a distinct --node-rank. --ngpu means GPUs per node; allocate nodes with your scheduler before launching.
• Data: there is no local/data.sh or db.sh, and the stages above do not prepare anything. Supply prepared manifests and length statistics directly. The local/, scripts, pyscripts, steps, and cluster configuration files used by the traditional pipeline are therefore omitted.

The default configuration is conf/train.yaml; additional stages live in conf/tuning/. See each recipe's README for stage selection and checkpoint usage.