CyanBridge speaker-verification exports

LiteRT exports used by CyanBridge's "Answer only my voice". Each graph takes 2 s of 16 kHz mono PCM in [-1, 1] (float32[1, 32000]) with the filterbank inside the graph and returns one speaker embedding. The app compares embeddings by cosine similarity against the wearer's on-device template; no audio or embedding leaves the phone.

File Bytes SHA-256 Source License
ecapa_spkrec_voxceleb_2s.tflite 84,076,036 f88f074f4446310ea6cf5d4089ddfb13bb8b7a8de76437086e47886ba5883259 speechbrain/spkrec-ecapa-voxceleb @ 0f99f2d0, converted with litert-torch 0.9.4 Apache-2.0
wespeaker_resnet34_lm_2s_int8w.tflite 11,632,752 56afd34101f2ffd9f6b016daaf6b8538576712edf2ba78ff096e18a7c1b4c53a Wespeaker/wespeaker-voxceleb-resnet34-LM @ f0c48c29, converted with litert-torch 0.9.4, ResNet weights int8 (weight-only) CC BY 4.0

Attribution

  • ECAPA-TDNN: SpeechBrain, speechbrain/spkrec-ecapa-voxceleb, Apache License 2.0 (https://www.apache.org/licenses/LICENSE-2.0). Modified: exported to LiteRT with the STFT as a fixed convolution; weights unchanged.
  • WeSpeaker ResNet34-LM: WeNet community, Wespeaker/wespeaker-voxceleb-resnet34-LM, Creative Commons Attribution 4.0 International (https://creativecommons.org/licenses/by/4.0/). Modified: exported to LiteRT with the Kaldi filterbank inside the graph; ResNet weights quantized to int8 (weight-only).

Both were trained on VoxCeleb by their authors. No CyanBridge user data was used.

Downloads last month
18
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support