Pseudonaut/chatterbox-nano-de-lora

Chatterbox Nano DE LoRA

Experimental German LoRA adaptation for ResembleAI Chatterbox Nano.

This repository contains adapter weights only. It does not redistribute the Chatterbox Nano base-model weights or the training dataset.

The adapter was trained with the public Chatterbox Nano Language Training pipeline.

Status

This is an experimental v0.1 release intended for research, local experimentation, and community testing.

German speech is generally intelligible and runtime speaker conditioning has been tested successfully with multiple reference voices. Occasional pronunciation, prosody, or generation errors can still occur. Quality may vary with text, punctuation, sentence length, and the reference voice.

This release should not be described as production-ready or error-free.

Base model

• Model: ResembleAI/chatterbox-nano
• Pinned model revision: 71ccd1d0081b430592cea481f4307e764e07bc64
• Chatterbox source revision used by the training environment: 5de7a54aa4e5e2baadb0182dde554908b48b85c2
• Nano checkpoint: t3_nano_v1.safetensors
• Backbone: GPT2_small

The base model is distributed separately by ResembleAI under the MIT license.

Training data

German reference dataset:

• Dataset: MrDragonFox/DE_Emilia_Yodas_680h
• Dataset revision: 4f70595a915782f741613dbad287bb830df21217
• Dataset license: CC BY 4.0
• Selected training target: approximately 20 hours
• The complete 680-hour dataset was not downloaded for this experiment.

The separately published raw-timestamp companion dataset was not used and is not included in this repository.

Training configuration

The published adapter comes from the 20-hour V2 scaling experiment.

• Strategy: LoRA
• Maximum run length: 10,000 optimizer updates
• Effective batch size: 16
  • micro-batch size: 1
  • gradient accumulation: 16
• Precision: bf16
• Learning rate: 1e-4
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Target modules:
  • c_attn
  • c_proj
  • c_fc
• Text embeddings trained: no
• Speaker encoder trained: no
• Gradient checkpointing: enabled
• Maximum text tokens: 256
• Maximum speech tokens: 850

The exact reference configuration is included as training_config.yaml.

Selected checkpoint

The released adapter is the best checkpoint selected from the 10,000-step run:

• Step: 9,500
• Recorded validation loss: 4.077134609

The validation set used by this experiment is limited, so the small numerical differences between nearby checkpoints should not be interpreted as a broad benchmark of German TTS quality.

The adapter file is approximately 9 MiB and contains PEFT LoRA weights only.

Speaker conditioning

Runtime speaker conditioning remains functional after adaptation.

The German adapter has been tested successfully using multiple runtime reference voices. This supports that the adaptation is not tied to one particular reference speaker.

This does not imply that every possible reference recording, speaker, microphone, accent, or acoustic condition has been validated.

Accent behavior

Accent characteristics depend strongly on the runtime reference voice.

In our tests, an English-accented reference voice transferred an audible English accent into otherwise German output. When a native German speaker was used as the reference voice, this English accent was not present.

For best German pronunciation, a clean native-German reference recording is therefore recommended.

This behavior also suggests that runtime speaker conditioning preserves more than speaker timbre alone and can carry accent and pronunciation characteristics from the reference recording.

Usage

This adapter is not a standalone TTS model. The Chatterbox Nano base model must be loaded separately and the LoRA adapter applied to its T3 transformer.

The reference implementation and evaluation tooling are available in:

https://github.com/Pseud0naut/chatterbox-nano-language-training

With that repository checked out and its environment prepared, an adapter can be evaluated with a runtime voice reference using the project's evaluation command. Point --adapter at the downloaded adapter directory and --voice-reference at a WAV file.

Example:

.venv/bin/python -m nano_de.evaluate \
  --device cuda \
  --adapter /path/to/chatterbox-nano-de-lora \
  --output-dir outputs/hf-adapter-test \
  --voice-reference /path/to/reference.wav

The training repository pins the upstream code and model revisions used for the reference result.

Known limitations

• Occasional German pronunciation errors remain.
• Prosody can be imperfect or unstable for some inputs.
• Longer or unusual sentences may produce generation errors or degraded phrasing.
• Pronunciation and accent characteristics can be influenced by the runtime reference recording; native-German reference speech is recommended for the best German pronunciation.
• The tokenizer was not expanded specifically for German.
• The adapter has not been exhaustively evaluated across accents, dialects, speakers, microphones, or domains.
• The reported validation loss comes from the project's limited validation setup and should not be treated as a standardized external benchmark.
• This is a research/community release, not a production guarantee.

Community reports of difficult words, sentence patterns, regressions, and reference-voice failures are useful for future versions.

Reproducibility

Training, preprocessing, diagnostics, evaluation code, pinned revisions, and reference configurations are published here:

https://github.com/Pseud0naut/chatterbox-nano-language-training

The public repository also documents the earlier invalid V1 training contract, the corrected V2 causal framing, semantic smoke tests, and later scaling experiments.

Licensing and attribution

The adapter files in this repository are released under the MIT License.

The upstream Chatterbox Nano model is a separate work distributed by ResembleAI under its own MIT license and is not redistributed here.

The German training dataset MrDragonFox/DE_Emilia_Yodas_680h is a separate work distributed under CC BY 4.0 and is not redistributed here. Please retain the dataset attribution when describing or redistributing results derived from this training setup.

See NOTICE for provenance and attribution details.

Acknowledgements

• ResembleAI for Chatterbox and Chatterbox Nano.
• MrDragonFox and the upstream Emilia/YODAS contributors for the German speech dataset used by the reference training run.