ImportError: cannot import name 'GenericForTokenClassification'
You installed Stage 5 into the Stage 1-4 environment. The vendored Chatterbox
package pins transformers==4.46.3; Stage 3 needs >=4.53. Use the two
environments described in Setup.
ERROR: Package 'chatterbox-tts' requires a different Python
Stage 5 is Python 3.11 only. Create .venv-tts with python3.11.
[WARN] No input JSONs / Number of human annotations: 0
Your --input_root points at the released Hugging Face layout, which is not
what the stages read. Convert it first with tools/unpack_corpus.py — see
CORPUS.md.
ImportError: FlashAttention2 has been toggled on, but it cannot be used
You passed --attn_implementation flash_attention_2 without the optional
flash-attn package. Install it (.venv/bin/pip install flash-attn --no-build-isolation, which compiles against your CUDA toolkit) or drop the
flag — the default auto falls back to PyTorch SDPA.
RuntimeError: All N requests failed from src.inference_turntaking_llm
Nothing answered at the endpoint, so every prediction would have been the
uniform fallback. Check that --base_url points at a running server, or that
OPENAI_API_KEY is exported if you passed an OpenAI model name. The message
quotes the first underlying error.
No input dialogues under <path> for split 'train' from src.synthesis.run
--input_root must be the parent of text_dialogue_<dataset>/, and Stage 4
reads the dialogues/ half of the corpus, not annotations/. Unpack it with
tools/unpack_corpus.py --kind dialogues, whose default root is
data-dialogues/. Note the released dialogues/ half is train-only.
ModuleNotFoundError: No module named 'src'
Run from the repository root, or .venv/bin/pip install -e ..
Building pynini fails
It needs OpenFst headers. sudo apt-get install libfst-dev, or install
pynini from conda-forge.
No input dialogues matched --input_glob from tts_render/convert_spoken.py
Stage 5 reads Stage 4's output, <save_root>/text_dialogue_<dataset>/<split>/ *.json. Quote the glob — an unquoted **/*.json is expanded by the shell
before Python sees it, and most shells do not expand ** recursively by
default.
--num_variants N needs that many distinct user voices
The bundled tts_render/librispeech_samples/ holds 12 speakers, and each
variant of a dialogue takes a different one. Lower --num_variants or point
--librispeech_root at a full LibriSpeech download.
Rendered backchannels sound flat or mispronounced Expected, and not something a flag fixes. Chatterbox is weak on two-phoneme interjections; the paper's dialogues used the ElevenLabs API for backchannels instead. See the note in Stage 5 for what to replace.