Vosk Russian TTS 0.9 Multi β€” original archive and BERT-quantized browser copy

Third-party model, redistributed with an explicitly modified copy. The Vosk TTS model and its voices originate from Alpha Cephei; this repository does not claim to have trained or authored them. The model-quant.zip file was made by quantizing one ONNX submodel for browser delivery, not by creating a new speech model.

Original source and authorship

Archives and published integrity metadata

File Role Size Published SHA-256
vosk-model-tts-ru-0.9-multi.zip Rehosted original-size archive 782,787,154 bytes 0aa332451ce46bfdbd620e74765fc16a4087988067c299969121ed0f8ed5bdf2
model-quant.zip Modified, BERT-quantized browser distribution 280,394,098 bytes 34e742ce9bb3c1ae86679d5974d2496b9fae50f0629f51bb4f5edfadc5ff3d71

The original-size archive's Hugging Face LFS SHA-256 equals the GitHub release asset's published SHA-256 in FedorMilovanov/gb-vosk-tts release model-v1. The quantized archive's Hugging Face LFS SHA-256 equals the expected download hash pinned by src/vosk-tts-engine.js. These are published provider metadata comparisons, not freshly computed hashes of independently downloaded ZIPs. Equal-size original files must not be treated as independently proven copies of the official Alpha Cephei archive until its publisher MD5 and SHA-256 are checked against downloaded bytes.

Modifications to the quantized archive

The project's archived July 2026 integration report describes the following changes and tests:

  • ONNX Runtime quantize_dynamic() was used for dynamic INT8 weight-only quantization of the BERT ONNX submodel, reducing its reported size from approximately 654 MB to 156.5 MB.
  • The main VITS synthesis model was not shipped in quantized form: its attempted quantization produced an inference-time MatMulInteger shape failure, so the original model was retained.
  • The full model ZIP was reduced from approximately 782 MB to 280 MB. Real speech synthesis/listening was reported as working in that historical test.
  • This archive is a modified redistribution and must not be represented as an unmodified Alpha Cephei model or as an independently trained model.

Reproducibility boundary: the archived record does not specify an exact command line, ONNX Runtime/Python versions, source and output ONNX hashes, or a fully reproducible per-file manifest. Historical listening results are not equivalent to an independently repeated evaluation. Byte-for-byte reproducibility of model-quant.zip has not been established.

Intended use

Russian multi-speaker speech synthesis via the external Vosk TTS inference pipeline and ONNX runtime. The project-specific browser integration is separate from these third-party model weights. Quality, compatibility and privacy depend on the application and its delivery path; this model card does not promise a particular browser, performance level, or production availability.

Rights and notices β€” important limitation

  • Original model/code credit: Alpha Cephei and contributors to Vosk TTS. The separate model repository does not transfer ownership of the model or its voices to the uploader.
  • Upstream license reference: Apache License 2.0. The current repository metadata declares apache-2.0, but a metadata label by itself is not an independent audit of the original model weights or their included notices.
  • Modification notice: model-quant.zip includes a BERT ONNX submodel changed via dynamic INT8 quantization; the original VITS synthesis submodel was retained.
  • Outstanding distribution checks: retrieve the official Alpha Cephei ZIP and verify its MD5 against the publisher's registry; independently compute SHA-256 for the official original, the GitHub release, the rehosted original and the modified ZIP; inspect archives for original license/NOTICE files and preserve all applicable notices; document exact quantization environment, transformation fingerprints and output audit. If the original archive contains an upstream NOTICE file, preserve its applicable content with any modifications as required by the license.

Verification status: publisher attribution and hosted-file metadata chain documented; independent upstream ZIP identity and complete modified-model redistribution clearance not yet established. Do not cite this card as proof of final rights clearance.

Engineering and provenance records

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support