Fixed Hub revisions and offline reproduction

Revision resolution preserves an explicit full 40-character lowercase Git commit without a separate model_info request. Branches, tags, abbreviated commits and unspecified revisions are resolved online. Offline use of such mutable references fails with an instruction to pin a full commit first. This applies to Qwen3-TTS, F5-TTS checkpoints/vocoders, PlayDiffusion presets and OzSpeech's downloaded codec assets, MaskGCT's primary and semantic checkpoints, XTTS-v2 and ZipVoice. Snapshot downloads may still contact the Hub when online.

The pin identifies a version; it does not prove that the required files exist in the local cache. A missing file still fails during snapshot or model loading. The Hub's download documentation describes full commit revisions and versioned snapshots.

For Qwen3-TTS Hub checkpoints, the runner resolves the pinned snapshot to a local directory before loading the upstream wrapper. The installed upstream wrapper forwards loading options to its model, but loads its processor separately without forwarding revision. Using the local snapshot keeps both components on the selected version and avoids the processor's repo metadata request in offline mode. No shared upstream package is modified.

The runner records the Hub repo and commit under model_reference.assets, alongside hashes of supported asset files in the snapshot. Its effective adversary.checkpoint_path is the local snapshot directory; the original configured repo remains in the saved run configuration. Explicit local checkpoints continue to use their actual file hashes.

With the model already cached, a fixed LibriTTS subset can be generated using:

HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 python run_vc.py \
  --config-name ots_vc/clean/libritts/qwen3_tts_ots \
  dataset.use_hf_dataset=false \
  +dataset.manifest_filename=/absolute/path/to/reproduction/subsets/libritts16_v1/metadata.json \
  adversary.revision=fd4b254389122332181a7c3db7f27e918eec64e3 \
  adversary.max_samples=16 +vc.generate_only=true +seed=42

This is an example revision used in the retained reproduction artifacts, not a claim that every upstream revision supports the same cloning interface.

F5-TTS assets

F5-TTS resolves checkpoint and vocoder revisions independently. The runner supplies local versioned assets to the upstream API, records their hashes, and hashes the installed preset configuration and default vocabulary. Explicit ckpt_file, vocab_file and vocoder_local_path overrides are retained. Checkpoint paths preserve the logical .safetensors filename in Hub caches; resolving the symlink to a bare blob filename can select the wrong loader.

With both snapshots cached, the validated default preset can run offline:

HF_HUB_OFFLINE=1 python run_vc.py \
  --config-name ots_vc/clean/libritts/f5_tts_ots \
  dataset.use_hf_dataset=false \
  +dataset.manifest_filename=reproduction/subsets/libritts16_v1/metadata.json \
  adversary.revision=84e5a410d9cead4de2f847e7c9369a6440bdfaca \
  adversary.vocoder_revision=0feb3fdd929bcd6649e0e7c5a688cf7dd012ef21 \
  +vc.generate_only=true +seed=42

These pins apply to F5TTS_v1_Base and its Vocos vocoder. Another preset or vocoder needs compatible revisions or explicit local assets. Current asset provenance does not reconstruct missing checkpoint identities from legacy runs.

MaskGCT checkpoints

The runner resolves adversary.repo_id and adversary.revision into an explicit adversary.checkpoint_dir containing all six native checkpoint files. It records their hashes before starting the worker; missing or empty files fail asset resolution. The worker loads these local paths without downloading replacement primary weights. Explicit local checkpoint directories bypass Hub resolution.

Offline resolution requires a full commit and all six files in the cache. The cached revision 265c6cef07625665d0c28d2faafb1415562379dc has been resolved and hashed locally. This verifies asset availability, not native generation after the backend migration or equivalence to historical weights.

The runner also resolves adversary.semantic_revision for facebook/w2v-bert-2.0 to adversary.semantic_model_path, containing its model configuration, feature-extractor configuration and safetensors weights. Both upstream semantic loaders use this same snapshot with local_files_only=True. The redirection is scoped to startup inside the isolated worker and restores the original loading methods afterward. Upstream files are not modified. Explicit local semantic snapshots bypass Hub resolution and must contain all three required files.

The cached semantic revision da985ba0987f70aaeb84a80f2851cfac8c697a7b has been hashed and loaded with the native SDK on CPU without missing, unexpected or mismatched weights. Its feature extractor matches the currently cached default processor on all 16 frozen references. These checks do not establish complete MaskGCT generation after migration or historical weight equivalence.

For offline runs, provide both full revisions after caching their required files:

HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 python run_vc.py \
  --config-name ots_vc/clean/libritts/maskgct_ots \
  dataset.use_hf_dataset=false \
  +dataset.manifest_filename=reproduction/subsets/libritts16_v1/metadata.json \
  adversary.revision=265c6cef07625665d0c28d2faafb1415562379dc \
  adversary.semantic_revision=da985ba0987f70aaeb84a80f2851cfac8c697a7b \
  +vc.generate_only=true +seed=42

The local dataset root must contain the referenced audio. Configure the MaskGCT worker environment through MASKGCT_RUNTIME_PYTHON or adversary.runtime_python before running. The runner hashes local semantic normalization statistics when code_path is configured. It also hashes the G2P source/data files, including the eagerly imported Chinese ONNX model and dictionaries. An eSpeak probe uses the worker interpreter and native-library search path; its version, selected shared library and complete data directory are recorded, including compiled dictionary files without extensions. Before model loading, the worker recomputes the eSpeak binary/data hashes and checks them against the resolved configuration. The parent also checks the startup receipt; changed resources or a missing receipt reject startup and reap the worker. Native English tokenization has been checked on the 32 prompt/target texts in the frozen 16-pair subset. Other-language package resources and a clean environment lock remain incomplete; these checks do not establish full generation or historical equivalence.

XTTS-v2 checkpoints

The runner resolves adversary.checkpoint and adversary.revision to a local snapshot before creating XTTS. It validates and hashes the native model, JSON configuration and vocabulary; a preset speaker bank is also captured when present. The native cloning path supports checkpoints without a preset speaker bank. Explicit file overrides are validated, hashed and passed to the loader. Missing or empty required files fail before loading weights.

The cached revision below resolves offline, and its four consumed file hashes match the retained LibriTTS16 run. Migrated native generation and scoring remain pending; these asset checks do not establish output equivalence or a complete runtime environment lock.

HF_HUB_OFFLINE=1 python run_vc.py \
  --config-name ots_vc/clean/libritts/xtts_ots \
  dataset.use_hf_dataset=false \
  +dataset.manifest_filename=reproduction/subsets/libritts16_v1/metadata.json \
  adversary.revision=6c2b0d75eae4b7047358e3b6bd9325f857d43f77 \
  adversary.local_files_only=true +vc.generate_only=true +seed=42

Standalone XttsGeneratorConfig accepts the same revision and local_files_only options. Explicit local checkpoint directories bypass Hub resolution.

ZipVoice assets

The runner resolves the ZipVoice model and Vocos vocoder independently, using adversary.revision and adversary.vocoder_revision. The selected zipvoice or zipvoice_distill directory contains the configured checkpoint, model.json and tokens.txt. The vocoder contains config.yaml and pytorch_model.bin. Missing or empty required files fail before loading models. Their local paths and hashes are recorded, and both native and CLI inference receive those explicit directories. Local model_dir, checkpoint_name and vocoder_path overrides are retained.

The following cached model and vocoder snapshots resolve offline; all five consumed file hashes match the retained Emilia-tokenizer LibriTTS16 run. This does not establish migrated native generation equivalence or freeze all tokenizer/package resources. The model pin below has been checked for zipvoice; another model variant needs its own compatible cached files.

HF_HUB_OFFLINE=1 python run_vc.py \
  --config-name ots_vc/clean/libritts/zipvoice_ots \
  dataset.use_hf_dataset=false \
  +dataset.manifest_filename=reproduction/subsets/libritts16_v1/metadata.json \
  adversary.revision=4ed45fb6e7e9527b780bef9e097a04bf13fe4e6b \
  adversary.vocoder_revision=0feb3fdd929bcd6649e0e7c5a688cf7dd012ef21 \
  adversary.local_files_only=true +vc.generate_only=true +seed=42

Provide the native ZipVoice source via adversary.code_path. Pin resolution applies to the benchmark runner. Standalone ZipVoiceGenerator callers need explicit local model_dir and vocoder_path to avoid implicit upstream downloads; both runtimes recheck configured local required files before model startup.

SparkTTS local assets

SparkTTS's relative model_dir resolves under its code_path, matching the native loader even if a different directory with the same name exists in the current working directory. The runner records the resolved absolute directory and hashes its assets. Before native imports, both runner and generator check configuration, BiCodec weights, LLM weights/tokenizer and wav2vec2 weights/feature-extractor files. Standard single-file or indexed Hugging Face weight layouts are accepted; every referenced shard must be present and nonempty.

Native BiCodec loading rejects missing learned state, unexpected checkpoint entries and shape mismatches. Only the two configuration-derived registered mel buffers, mel_transformer.spectrogram.window and mel_transformer.mel_scale.fb, may be absent. A missing entry is accepted only as a registered non-trainable finite buffer. Their shapes, dtypes and value hashes are recorded in each successful sample's native_initialization receipt and retained during reuse. The scoped check restores the native class method after startup or failure; upstream source files remain unchanged.

The current local bundle's 14 asset hashes match the retained LibriTTS16 run. Actual native BiCodec CPU loading accepts its 840 checkpoint entries with only those two derived buffers absent, and rejects removal of a learned encoder tensor in memory. This component check used the existing audiobench runtime; the default controller environment lacks einx. It does not establish full SparkTTS generation after migration, LLM/wav2vec loading completeness, a clean runtime lock or immutable historical weights. Use a compatible SparkTTS runtime with a complete explicit local model bundle for full generation.

CosyVoice local assets

CosyVoice requires an explicit local model_dir. Discovery stays inside that directory, and the runner binds and hashes the effective payload directory. Runner and generator reject missing or empty core checkpoints, the selected variant's YAML configuration and speech tokenizer before importing the native SDK. A configured zero_shot_spk_id also requires a nonempty spk2info.pt bank; its actual speaker membership is still checked by native inference.

CosyVoice2's native constructor overrides its Qwen pretrained path to CosyVoice-BlankEN under the payload directory. The preflight therefore also requires its configuration, vocabulary, merges and weights. Standard single-file or indexed Transformers weights are accepted, with every declared shard present and nonempty. The generator rechecks these files at model loading. CosyVoice1 receives only the arguments accepted by its constructor; load_vllm is restricted to CosyVoice2.

These checks validate local file completeness, not tensor loading or speech quality. CosyVoice1's YAML-selected text assets, custom YAML dependencies, JIT/TensorRT/vLLM assets and text-frontend resources need further runtime evidence. The previously measured native subset uses CosyVoice2 with accelerators disabled; post-migration native subset generation remains pending.

OpenVoice converter loading

OpenVoice's native converter uses strict=False. The wrapper requires a complete finite checkpoint with matching tensor shapes instead. Missing or unexpected keys, shape errors and nonfinite tensors fail startup. The check is scoped to the converter's owned Torch model; other Torch modules retain their behavior, and the original method is restored after success or failure. Startup must perform exactly one checked load. Successful rows record the converter's key count, strictness and compatibility result in native_initialization, retained on resume and saved-audio scoring.

Actual native converter CPU startup accepts the local 486-entry checkpoint with no missing or unexpected state. Its 32,792,226 parameters load without changing any checkpoint tensor, and removing a learned decoder parameter in memory is rejected. This check used the existing audiobench environment with a writable Numba cache and bypassed the Melo import stage to isolate the converter.

This verifies the converter component only. Melo checkpoints, BERT/text assets, speaker embeddings and post-migration native subset generation still require separate evidence. A compatible full OpenVoice/Melo runtime is required; loading the standalone converter architecture does not validate its audio API imports.

OpenVoice Melo language models

The LibriTTS OpenVoice configuration pins the English Melo model to myshell-ai/MeloTTS-English revision bb4fb7346d566d277ba8c8c7dbfdf6786139b8ef. The runner resolves its configuration and checkpoint from that same snapshot, checks both files are nonempty, records their hashes, and supplies their explicit local paths to the native constructor. adversary.local_files_only=true uses an already cached pinned snapshot.

Configure additional uppercase language entries in adversary.melo_models with repo_id and revision, or both config_path and checkpoint_path for explicit local files. Once this mapping is configured, an absent language fails rather than falling back to an implicit download. Standalone generator calls require resolved local paths in the mapping. Older configurations without a mapping retain the native implicit asset behavior as a separate unpinned path.

The actual pinned English architecture accepts all 1,051 checkpoint entries with no missing or unexpected state; its 51,874,097 parameters match every checkpoint tensor and all tensors are finite. This CPU component check excludes Melo's text API and speech inference. The earlier native subset did not record Melo asset hashes, so it does not establish historical checkpoint identity. Transformer text assets use the separate text_models mapping described below. Full native text and speech validation remains pending.

OpenVoice English pronunciation resources

The LibriTTS configuration enables adversary.capture_english_text_resources=true. Before model loading, the runner records Melo's CMU dictionary and its existing pickle cache, the trained g2p_en parameter archive, and NLTK's CMU/tagger resources. It includes both archives checked by g2p_en during import and the active corpus/tagger used by the installed NLTK version. Directory resources include extensionless files. Missing resources fail preflight; this path does not call a downloader.

These are resources discovered in the generation interpreter. Configure melo_code_path for the intended checkout and install the needed NLTK data in that runtime before running. Matching hashes identify the observed resource files; they do not pin their installation source or guarantee a clean runtime. The local inventory records seven resource entries and ten files. BERT and its tokenizers, non-English resources and migrated speech generation remain outside this check.

OpenVoice Transformer text assets

The LibriTTS configuration lists fixed snapshots in adversary.text_models, keyed by the upstream repository IDs used by Melo. Melo eagerly imports six language tokenizers even for English generation. The mapping provides these tokenizers and enables weights only for English bert-base-uncased with load_model: true. The runner validates and hashes the selected files; cached weights for other languages are not included unless enabled.

During native imports and text inference, scoped loading redirects configured repository IDs to their resolved local snapshot paths with local_files_only and remote custom code disabled. Unlisted IDs and weight loading on a tokenizer-only entry fail explicitly. Loader methods are restored after success or failure. Configurations without text_models retain the separate legacy implicit-loading path. Standalone generator calls require resolved local paths.

This fixes asset selection, not Transformer version compatibility or all text resources. The observed runtime also emits a tokenizer regex warning for the multilingual snapshot. Actual native English G2P, phone/tone/language tensors and expanded BERT features have repeated identically for the 16 selected target texts on CPU. Historical repo-ID tokenization equivalence and speech reproduction remain unproven.

OpenVoice tracks Melo text module objects imported by its own serial operations. On close, it releases their native global Torch model caches as well as its language models and reference embeddings. A replacement module or a previously imported foreign module is not selected for cleanup. Partial import failures still record owned modules for cleanup. The native LibriTTS16 text check verifies that closing the generator releases the actual English BERT cache without changing any of the checked text feature hashes.