Releases: VoxRT/voxrt-asr-models
Releases · VoxRT/voxrt-asr-models
Release list
v0.1.2 — streaming-medium-pc weights (re-tagged for ASR iOS v0.1.2)
Re-tag of v0.1.1 — the .vxrt file is bit-identical to v0.1.1 (SHA-256:
0d723e429157a8a8cb58739a1f090574f2f23db311ca7916b43411f5f727c79c).
This tag exists so the iOS voxrt-asr-ios v0.1.2 consumer README links resolve.
Pair with:
- iOS:
voxrt-asr-iosv0.1.2 (xcframework Headers/ collision fix) - Android:
voxrt-asr-androidv0.1.1 (unchanged from v0.1.1 release)
Download
curl -fLO https://github.com/VoxRT/voxrt-asr-models/releases/download/v0.1.2/streaming_medium_pc.vxrtSHA-256: 0d723e429157a8a8cb58739a1f090574f2f23db311ca7916b43411f5f727c79c
v0.1.1 — streaming-medium-pc weights
First successful public release of the streaming-medium-pc ASR model weights, packaged for the VoxRT on-device inference runtime.
(v0.1.0 was tagged but its companion voxrt-asr-android library AAR failed to build on JitPack; v0.1.1 is the fixed version. Use v0.1.1
throughout.)
Pair with one of the consumer libraries:
- Android —
voxrt-asr-android - iOS — coming soon
What's inside
streaming_medium_pc.vxrt— NVIDIA NeMo FastConformer Hybrid Medium (32M parameters), packed into VoxRT's encrypted.vxrt v2format. ~61
MB on disk. Bit-identical to v0.1.0 — only the version tag differs.- Hybrid CTC + RNN-T decoder heads. Cache-aware streaming with ~1.1 s chunks.
- WER on LibriSpeech test-clean (500 utterances): 3.267 % (RNN-T) / 4.895 % (CTC).
Download
curl -fLO https://github.com/VoxRT/voxrt-asr-models/releases/download/v0.1.1/streaming_medium_pc.vxrtSHA-256: 0d723e429157a8a8cb58739a1f090574f2f23db311ca7916b43411f5f727c79c
Use it
implementation("com.github.VoxRT:voxrt-asr-android:v0.1.1")License
- Weights derived from
nvidia/stt_en_fastconformer_hybrid_medium_streaming_80ms_pc, released under CC-BY-4.0. .vxrtformat + encryption are VoxRT-specific runtime artefacts.