teamy-transcriber-whisper-large-v3

Teamy Transcriber: ready Whisper large-v3 package

Data-only model package for the source-defined Rust/CUDA runtime in Teamy Transcriber. The native CLI supporting this package provides transcribe and local model selection. Use that native build; earlier application builds may not support these commands. This repository contains no inference code, TorchScript graph, Python module or audio recordings. The model computation belongs in the application's source code.

Download and select

hf download TeamDman/teamy-transcriber-whisper-large-v3 --revision v1 --quiet
teamy-transcriber model prepare
teamy-transcriber transcribe "example.webm"

hf download --quiet prints the snapshot directory. For explicit selection:

$model = hf download TeamDman/teamy-transcriber-whisper-large-v3 --revision v1 --quiet
teamy-transcriber model prepare --source-dir $model

No-argument preparation discovers the release in the local HF cache, validates it and saves the selected folder in application settings. It does not download, copy or re-encode these prepared weights. Keep the cached snapshot available; if you remove it, download and select it again. An explicit TEAMY_TRANSCRIBER_MODEL_DIR environment variable overrides saved selection.

Contents and provenance

• Original canonical Whisper large-v3 safetensors, config, tokenizer and greedy generation policy from openai/whisper-large-v3 at revision 06f233fe06e710322aca913c1bc4249a0d71fce1.
• dims.json: small derived dimension metadata consumed by Teamy Transcriber.
• vad/: prepared Silero 16 kHz speech-detector tensors, schema/checksum manifest and original MIT license. Detection runs as source-defined CPU code.
• model-manifest.json: asset sizes, SHA-256 checksums and provenance.

Whisper weights occupy 3,087,130,976 bytes (about 3.09 GB decimal). They are unchanged, not newly trained or quantized. The application currently expands weights to FP32 GPU storage, so VRAM use is larger than the downloaded file.

Compatibility and limitations

The native runtime requires an NVIDIA GPU and CUDA runtime libraries. Validation is on Windows x64 / RTX 4090, English greedy transcription, with clip timestamps and optional speech detection. This package is not a promise of CPU Whisper, multilingual/translation support, word alignment, diarization or beam search in the current Teamy application. No hosted inference endpoint is provided.

Whisper can produce incorrect or hallucinated text, including on noisy or silent inputs; speech detection reduces silent input processing but does not guarantee accuracy. See the upstream model card for training information, intended use and limitations. This package adds no new training or evaluation claim for the underlying model.

Licenses

The upstream Hugging Face checkpoint declares Apache-2.0; its text is in LICENSE. Original OpenAI Whisper's MIT notice is preserved in LICENSE-OPENAI-MIT.txt. Silero's MIT license is in vad/LICENSE.txt. The added dimension and manifest metadata are distributed under Apache-2.0.