Final Publication Acceptance

GitHub Release and PyPI 1.4.25 are published. Both official PyPI distributions were downloaded and match GitHub by SHA-256 and size:

• funasr-1.4.25-py3-none-any.whl: 1015440 bytes; SHA-256 11ab7bd8c7574a94579b83a9f8fe837af46be6229462da5f3004a04a017cee75.
• funasr-1.4.25.tar.gz: 1083745 bytes; SHA-256 1ff96e278b7bf6823da0afa9290a12a84a0e641d31dd63fddaafd82a56000c80.

FunASR 1.4.25

Source: 455ad37cc03bd01e85740caf64788ee4cf8ebe2c / tag v1.4.25.

Fixes

• #3863: Python 3.12-compatible boolean argument parsing without distutils, retaining normalized invalid-value messages and boolean return types.
• #3866: remove synthetic leading spaces from Latin sentence timestamp output while preserving time boundaries, raw text and VAD alignment.

This root wheel does not publish the independent funasr-onnx runtime fix from #3865. Open SenseVoice #3861 and VAD #3862 changes are not included.

Verification Scope

• Corrected #3863 head: actual Python 3.10/3.12 hosted artifacts each have 53 passing tests, including 23 boolean-parser contracts. Corrected #3866 head: actual NumPy 1.26.4/2.4.0 artifacts each have 421 passes and 27 CUDA-only skips. Old failed runs are retained, not rerun.
• Combined source: 468 passes and 33 skips across the union of both CI selections. An earlier private run had one missing pytorch_wpe child-process import (467 passes); a process-only private dependency path corrected it without changing product source or assertions.
• Candidate release/docs checks: {'contract-green-v3': {'testcase': 49, 'failure': 0, 'error': 0, 'skipped': 0}, 'static-remainder': {'testcase': 104, 'failure': 0, 'error': 0, 'skipped': 2}, 'installed-behavior-candidate': {'testcase': 54, 'failure': 0, 'error': 0, 'skipped': 0}}. Actual installed candidate and final wheels exercise the parser, timestamps and mocked VAD integration; final counts: {'testcase': 54, 'failure': 0, 'error': 0, 'skipped': 0}. All 439 package files and 13 entrypoints match accepted source.
• Root model/GPU-runtime bytes are unchanged from 1.4.24 and remain bound to its accepted L20Z evidence. No new GPU execution, pretrained ASR-quality, full-suite or Nano result is claimed. The prior repository-wide offline attempt remains incomplete at online model lookup and timeout.
• Head, merged-main and final-tag hosted CI and actual checkout trees are separately accepted. Final Python wheel and sdist are newly downloaded and checked for metadata, exact source bytes, archive integrity, strict Twine, private offline installation and CLI behavior.
• Ten unchanged runtime archives reuse complete cache bytes, freshly rehashed against the new assets and original runtime release with archive-integrity checks. They were not rebuilt or re-executed. SHA256SUMS covers all 12 payloads.
• The 1.4.24 warning still applies: checkpoints trained with both norm_means=True and norm_vars=True need recognition-quality reevaluation. norm_vars=False remains the default.
• No production upgrade or adoption/growth claim.

What's Changed

• fix(onnx): load non-.wav audio paths with librosa instead of pydub by @tanishqbaweja in #3865
• fix(utils): remove distutils boolean parser by @GPT-AGI in #3863
• fix: drop the leading space on Latin sentence timestamps by @GPT-AGI in #3866
• Release FunASR 1.4.25 by @LauraGPT in #3875

New Contributors

• @GPT-AGI made their first contribution in #3863

Full Changelog: v1.4.24...v1.4.25

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.2.6.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.4.25"

FunASR ONNX Runtime 0.4.10

Install: pip install -U funasr-onnx==0.4.10.

This standalone package delivers #3865: offline/streaming Paraformer,
SenseVoice and offline VAD load supported non-WAV audio directly through
librosa, without an intermediate WAV sidecar or MP3-only conversion.
Codec support depends on the installed audio backend. Updating the root
funasr package does not update funasr-onnx.

Validation includes45 source contracts and42 actual installed tests for each
candidate wheel/sdist, including lossless FLAC, mixed WAV/FLAC lists, array
identity, invalid-audio handling and existing streaming/VAD/punctuation cases.
Official0.4.9 fails the two new FLAC/list tests through its pydub conversion path.
New exact-head/main Python3.11/3.12 CI verifies clean torch-free installation
and installed behavior; actual artifacts and checkout trees are checked.

Prior exact-source #3865 Paraformer L20Z/ORT-CUDA fixture evidence is retained
and reused:48000 samples, identical FLAC-path/array results,10 CUDA nodes/0 CPU
nodes. It is not new GPU execution or evidence of pretrained recognition
quality, all-model inference, BF16 or a full-root-suite pass. No new weights.
The open SenseVoice#3861 and VAD#3862 changes are not included.

Assets:wheel,sdist,SHA256SUMS covering two payloads. Final bytes come from the
accepted merged-main CI. Runtime tag/source and all assets are checked; no
final-tag CI is claimed when current workflows do not select this runtime tag.
This is a non-latest runtime release and preserves root1.4.25 as latest.
PyPI upload is last and is accepted only after official wheel/sdist downloads
match GitHub SHA256 and size. No Nano, production upgrade or growth claim.

Final Publication Acceptance

GitHub Release and PyPI 1.4.24 are both published. Both distributions were downloaded from official PyPI and match the accepted GitHub assets by SHA-256 and size:

• funasr-1.4.24-py3-none-any.whl: 1015161 bytes; SHA-256 c2678ebb77c7dfba1d85a79749ad30b441159b87d17eff901e0c6ad4ebcc06c7.
• funasr-1.4.24.tar.gz: 1081873 bytes; SHA-256 522dcb493be75aa57d8183e6197b4431c5e02fade689eda90cfd91f76e5edb57.

The official sdist download resumed after a transport timeout using an exact HTTP 206 range; the original partial and failure receipts were retained. The wheel was not downloaded again and PyPI was not re-uploaded.

FunASR 1.4.24

Source: 20f26b7c4905d010768ec91206149df2a90104b6 / tag v1.4.24.

Fix

Deliver #3852: ignore empty non-linguistic symbols instead of allowing the character/phoneme tokenizer scan to stall. Deliver #3853: correct variance normalization for padded utterances. Deliver #3854: preserve valid CIF output and gradients when a batch mixes utterances with zero and nonzero fires, including dynamic ONNX batches.

Compatibility: checkpoints trained with both norm_means=True and norm_vars=True require recognition-quality reevaluation after the corrected normalization. norm_vars=False remains the default. Local checkpoint round trips are not a pretrained recognition-quality evaluation.

This does not claim pretrained ASR quality, Windows/NPU acceptance or a full repository-suite result.

The separately released funasr-onnx==0.4.9 and funasr-torch==0.1.6 have their own tags, packages and verification. This root-package release does not republish them. PR #3848 is not included. This release does not claim every EOS path is fixed.

Verification Scope

• New candidate release/README checks: 45 cases, 0 skips. Related docs/package/release checks: 104 cases, 2 explicitly recorded skips; no failures or errors.
• Actual installed candidate and final wheels: 37 passing CPU/ONNX cases each covering tokenizer symbols, variance normalization, local checkpoint round trips, repository-audio inference and CIF dynamic batches. The 27 CUDA parametrizations are intentionally skipped locally and separately covered by bound prior hardware evidence. All 439 package files and 13 entrypoints match the accepted source.
• Reused 24 tokenizer/CTC cases, 34 FP32 CIF cases and four normalizer repository-audio cases including six CUDA inference calls on L20Z, after rehashing 33,870 sealed evidence files and 18 helpers and binding the exact source to the installed wheel. The 106-case normalizer acceptance retains 102 unchanged cases and four repaired audio checks. Earlier harness/TF32-fixture failures remain preserved; they are not silently discarded or counted as passes. This release performs no new GPU execution and makes no pretrained-quality or full-Paraformer-export claim.
• The local candidate build and candidate/final package checks, tests and installations use private HOME/cache directories and separate network namespaces with verified network-unreachable probes. The final-tag build runs in GitHub Actions under the existing release workflow, not the local network namespace. No newly downloaded pretrained weights are used. The separately reported #3848 test-download incident is not an authorization or validation input for this release.
• Release-head CI was accepted from actual logs and recorded artifact inventories. The roadmap job checked out the pure head; the other head workflows used GitHub's temporary PR merge commit. Both checkout trees were verified against the exact candidate. Merged-main and final-tag source checkouts were separately verified. The main API-doc workflow also checked out only the retained Chinese API page from gh-pages into previous-pages; this separate documentation input was recorded, not treated as inference source. The newly downloaded Python wheel/sdist passed byte/metadata/archive/source checks, strict Twine, offline installation, actual import and CLI checks.
• The 10 unchanged runtime archives reuse complete cache bytes freshly hashed against new asset metadata and the existing runtime source release. They were not rebuilt or re-executed. SHA256SUMS covers all 12 payloads. This is not a new runtime-hardware claim.
• No full repository-suite, production-environment upgrade, Nano benchmark or adoption/growth claim is made.

What's Changed

• fix(tokenizer): prevent empty-symbol hangs with hosted regression coverage by @LauraGPT in #3855
• fix(tokenizer): ignore empty non-linguistic symbols to avoid a hang by @MohammadHijjawi97 in #3852
• fix(normalize): correct variance with reproducible CPU/CUDA regressions by @LauraGPT in #3856
• fix(normalize): correct UtteranceMVN variance normalization by @BillVan1945 in #3853
• fix(cif): handle mixed zero-fire batches with CPU/ONNX regression CI by @LauraGPT in #3857
• fix(paraformer): handle trailing zero-fire samples in mixed CIF batches by @BillVan1945 in #3854
• Release FunASR 1.4.24 by @LauraGPT in #3858

Full Changelog: v1.4.23...v1.4.24

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.2.6.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.4.24"

Final Publication Acceptance

GitHub Release and PyPI 1.4.23 are both published. Both distributions were downloaded from official PyPI and match the accepted GitHub assets by SHA-256 and size:

• funasr-1.4.23-py3-none-any.whl: 1014744 bytes; SHA-256 28f6696ad882d1273c371675ac35b427d2e548a2e736e42ef16d56cd96d17785.
• funasr-1.4.23.tar.gz: 1075334 bytes; SHA-256 89a2e1ba43468391bd603e952dd6fe306020719d82f99cd15bbdd2b130afd2f3.

FunASR 1.4.23

Source: efe5504e388613e5298e1f93d5ca9f861700b489 / tag v1.4.23.

Fix

Deliver #3849: initialize the valid-frame cursor for the online frontend with lfr_m=lfr_n=1, and avoid redundant empty final offline VAD chunks for exact input multiples. Deliver #3783: supplied token_list IDs map to model pieces before text conversion. With a supplied vocabulary, SentencePiece ids2tokens returns list[str] pieces; use tokens2text for text. Without a supplied vocabulary, legacy decoded-text behavior is preserved.

This does not claim pretrained ASR quality, Windows/NPU acceptance or a full repository-suite result.

The separately released funasr-onnx==0.4.9 and funasr-torch==0.1.6 have their own tags, packages and verification. This root-package release does not republish them. PR #3848 is not included. Explicit streaming-empty-input and nonempty tails producing zero features are outside the #3849 hardware acceptance; this release does not claim every EOS path is fixed.

Verification Scope

• New candidate release/README checks: 45 cases, 0 skips. Related docs/package/release checks: 104 cases, 2 explicitly recorded skips; no failures or errors.
• Actual installed candidate and final wheels: 61 CPU contracts each covering no-LFR feature/chunk handling, analytic VAD finalization and tokenizer vocabulary semantics. Local tiny SentencePiece and controlled acoustic fixtures are not pretrained quality or hardware inference. All 439 package files and 13 entrypoints match the accepted source.
• Reused 16 direct FP32/BF16 frontend cases, 12 FP32 VAD cases and 12 FP32/BF16 tokenizer/decoder cases from #3849/#3783 after rehashing their sealed raw receipts and binding the exact source to the installed wheel. The earlier frontend run's six failed VAD cases remain recorded; the later corrected VAD run passed all 12. This release performs no new GPU execution and makes no pretrained-quality claim.
• The local candidate build and candidate/final package checks, tests and installations use private HOME/cache directories and separate network namespaces with verified network-unreachable probes. The final-tag build runs in GitHub Actions under the existing release workflow, not the local network namespace. No newly downloaded pretrained weights are used. The separately reported #3848 test-download incident is not an authorization or validation input for this release.
• Exact release-head and merged-main CI were accepted from their actual logs, with artifact inventories recorded. The final-tag workflow and the newly downloaded Python wheel/sdist passed byte/metadata/archive/source checks, strict Twine, offline installation, actual import and CLI checks.
• The 10 unchanged runtime archives reuse complete cache bytes freshly hashed against new asset metadata and the existing runtime source release. They were not rebuilt or re-executed. SHA256SUMS covers all 12 payloads. This is not a new runtime-hardware claim.
• No full repository-suite, production-environment upgrade, Nano benchmark or adoption/growth claim is made.

What's Changed

• fix(frontend): handle default no-LFR online features by @LauraGPT in #3849
• fix(onnx): handle empty and short inputs in offline VAD by @BillVan1945 in #3846
• fix(tokenizer): respect the model vocabulary in ids2tokens by @carey-bk in #3783
• fix(onnx): handle empty punctuation input without consuming cache by @BillVan1945 in #3847
• fix(runtime): release ONNX VAD and punctuation fixes as 0.4.9 by @LauraGPT in #3850
• Release FunASR 1.4.23 by @LauraGPT in #3851

New Contributors

• @carey-bk made their first contribution in #3783

Full Changelog: v1.4.22...v1.4.23

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.2.6.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.4.23"

Standalone funasr-onnx 0.4.9. This is not a new root funasr release.

This release includes the already accepted offline VAD fix from #3846: empty or
sub-frame inputs preserve one result per input without invoking ONNX Runtime for
zero-frame features. It also includes #3847: empty punctuation chunks return
without tokenization or inference while preserving streaming cache state.

The hosted standalone-package workflow now runs the #3847 regression on Python
3.11 and 3.12. Exact-head and merged-main CI each passed 9 release contracts,
6 offline VAD cases, 5 empty-punctuation cases, the existing installed/runtime
regressions, Product site, browser, documentation, build, Twine, and clean
package checks. Local fail-closed private-network validation also passed 45
source/contract tests, 38 installed-wheel tests, and installed-sdist origin
verification.

These are CPU runtime-wrapper fixes. No model weights changed, no model was
downloaded or loaded for this release, and no ASR-quality claim is made.

Install: pip install -U funasr-onnx==0.4.9.

Assets: wheel, sdist, and SHA256SUMS covering those two payloads. PyPI is the
final upload step and is accepted only after downloading the official bytes.

Final Publication Acceptance

GitHub Release and PyPI 1.4.22 are both published. Both distributions were downloaded from official PyPI and match the accepted GitHub assets by SHA-256 and size:

• funasr-1.4.22-py3-none-any.whl: 1014515 bytes; SHA-256 83525bb5f185e29032f840f232eb226080ffa37e9d86988893faa225e609dc35.
• funasr-1.4.22.tar.gz: 1069356 bytes; SHA-256 db8c67f2bf36df25bd26e724d7a432ec04667e877cd81119d8d388a58ae82eb0.

FunASR 1.4.22

Source: dedefbf643289098e5473fa0cc6ce085e2e9533e / tag v1.4.22.

Fix

Deliver #3841: merge_length_s passed to generate() takes precedence over construction-time/default merge length. Deliver #3833: the dynamic silence schedule is evaluated per frame using the current speech-segment duration, not accumulated chunk time; the fixed-threshold opt-out remains available.

This does not claim pretrained ASR quality, Windows/NPU acceptance or a full repository-suite result.

The separately released funasr-onnx==0.4.8 and funasr-torch==0.1.6 have their own tags, packages and verification. This root-package release does not republish them. The separately reproduced no-LFR frontend exception is not fixed by this release.

Verification Scope

• New candidate release/README checks: 45 cases, 0 skips. Related docs/package/release checks: 104 cases, 2 explicitly recorded skips; no failures or errors.
• Actual installed candidate wheel: 9 CPU unit contracts covering per-call/construction/default merge length and dynamic/fixed VAD scheduling. Their mocked boundaries are not hardware inference. All 439 package files and 13 entrypoints match the accepted source.
• Reused the 32 AutoModel and 24 VAD real L20Z FP32 analytic integration cases from #3841/#3833 after rehashing their sealed raw receipts and binding the exact source to the installed wheel. Their separate prior runs record 216521 CUDA events in total. This is not a new GPU execution, BF16 run or pretrained-quality validation.
• Exact release-head and merged-main CI were accepted with actual logs/artifacts. The final-tag workflow and the newly downloaded Python wheel/sdist passed byte/metadata/archive/source checks, strict Twine, offline installation, actual import and CLI checks.
• The 10 unchanged runtime archives reuse complete cache bytes freshly hashed against new asset metadata and the existing runtime source release. They were not rebuilt or re-executed. SHA256SUMS covers all 12 payloads. This is not a new runtime-hardware claim.
• No full repository-suite, production-environment upgrade, Nano benchmark or adoption/growth claim is made.

What's Changed

• fix(runtime): decode every batch in ContextualParaformer and SeacoParaformer by @tanishqbaweja in #3838
• fix(runtime): release ONNX hotword batching fix as 0.4.7 by @LauraGPT in #3839
• fix(runtime): release Torch hotword batching fix as 0.1.5 by @LauraGPT in #3840
• fix(auto): honor merge_length_s passed to generate() by @MohammadHijjawi97 in #3841
• fix(vad): apply the offline dynamic silence schedule per frame by @lingfengyao in #3833
• fix(runtime): handle blank and incomplete CMVN files by @BillVan1945 in #3842
• fix(runtime): release ONNX CMVN parser fix as 0.4.8 by @LauraGPT in #3843
• fix(runtime): release Torch CMVN parser fix as 0.1.6 by @LauraGPT in #3844
• Release FunASR 1.4.22 by @LauraGPT in #3845

New Contributors

• @lingfengyao made their first contribution in #3833
• @BillVan1945 made their first contribution in #3842

Full Changelog: v1.4.21...v1.4.22

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.2.6.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.4.22"

Standalone Torch Runtime 0.1.6

Includes #3842: the standalone CMVN reader now accepts blank lines and raises a file-specific ValueError for missing or truncated statistics. Valid float64 statistics and shipped feature output remain unchanged.

python -m pip install -U "funasr-torch>=0.1.6"

funasr-torch is distributed separately from funasr. This is a CPU parser/package fix, not a model-weight or ASR-accuracy update.

The wheel and source distribution contain the accepted runtime source. Validation covers both standalone frontends, blank/truncated/missing/compact CMVN input, shipped am.mvn feature equality, package contracts, build/install metadata, and exact-head/main CI. It does not claim model inference, new GPU execution, Windows/NPU acceptance, or a full suite.

Assets: wheel, source distribution, and SHA256SUMS covering both payloads. GitHub asset bytes are verified before publication; PyPI upload and official payload-byte verification follow separately. The root FunASR latest release remains v1.4.21.

Standalone Torch Runtime 0.1.5

Includes the Contextual/Seaco Paraformer hotword batching and empty hotword-file handling fix from #3838.

python -m pip install -U "funasr-torch>=0.1.5"

funasr-torch is distributed separately from funasr. This runtime-wrapper release changes no model weights and makes no ASR accuracy claim.

Validation includes 14 real package tests, build/Twine/isolated installation checks, and exact-head CI. All 10 packaged runtime modules reuse the byte-identical #3838 source and its 52 selected CPU and 48 real L20Z FP32 analytic TorchScript/ONNX batching contracts across batch sizes 1/2/3. These are not new inference runs, pretrained ASR benchmarks, Seaco timestamp-output tests, Windows/NPU validation, or a full-suite result.

Assets: wheel, source distribution, and SHA256SUMS covering both payloads. GitHub bytes are verified before publication; PyPI upload and official downloaded-byte verification follow separately. The root FunASR latest release remains v1.4.21.

Standalone ONNX Runtime 0.4.8

Includes #3842: the standalone CMVN reader now accepts blank lines and raises a file-specific ValueError for missing or truncated statistics. Valid float64 statistics and shipped feature output remain unchanged.

python -m pip install -U "funasr-onnx>=0.4.8"

funasr-onnx is distributed separately from funasr. This is a CPU parser/package fix, not a model-weight or ASR-accuracy update.

The wheel and source distribution contain the accepted runtime source. Validation covers both standalone frontends, blank/truncated/missing/compact CMVN input, shipped am.mvn feature equality, release contracts, build/install metadata, and exact-head/main CI. It does not claim model inference, new GPU execution, Windows/NPU acceptance, or a full suite.

Assets: wheel, source distribution, and SHA256SUMS covering both payloads. GitHub asset bytes are verified before publication; PyPI upload and official payload-byte verification follow separately. The root FunASR latest release remains v1.4.21.

Standalone ONNX Runtime 0.4.7

Includes the Contextual/Seaco Paraformer hotword batching and empty hotword-file handling fix from #3838.

python -m pip install -U "funasr-onnx>=0.4.7"

funasr-onnx is distributed separately from funasr. This release changes runtime wrappers, not model weights, and does not claim an ASR accuracy improvement.

The wheel and source distribution contain the accepted runtime source. Validation includes release-contract/build/installed-package checks and exact-head CI. The unchanged inference modules reuse the existing #3838 evidence: 52 selected CPU tests and 48 real NVIDIA L20Z FP32 analytic TorchScript/ONNX batching contracts across batch sizes 1/2/3. This is not a new pretrained ASR benchmark, Seaco timestamp-output test, Windows/NPU validation, or full-suite result.

Assets: wheel, source distribution, and SHA256SUMS covering both payloads. GitHub asset bytes are verified before publication; PyPI upload and official payload-byte verification follow separately. The root FunASR latest release remains v1.4.21.