pat229988/speaker-embedding-w2v-bert-tflite-dynamic

Speaker Embedding W2V-BERT TFLite Package

This repository contains a two-stage TFLite/LiteRT speaker embedding package for local CPU testing: deterministic waveform feature extraction followed by a dynamic-range quantized W2V-BERT feature-to-embedding model.

Files

File	Purpose
speaker_feature_extractor.tflite	TFLite waveform-to-input_features preprocessing.
speaker_feature_extractor.tflite.manifest.json	Feature extractor conversion/runtime manifest.
speaker_features_to_embedding.tflite	TFLite input_features-to-embedding flatbuffer.
speaker_features_to_embedding.tflite.manifest.json	Conversion/runtime manifest.
artifact_manifest.json	Clean upload metadata, checksum, and source lineage.

Source lineage

This is a converted/quantized derivative of a speaker verification stack built from:

• Base model: facebook/w2v-bert-2.0
• Speaker checkpoint / adapter: zl389/w2v-bert-2.0_SV
• Original implementation repo: https://github.com/ZXHY-82/w2v-BERT-2.0_SV

Direct upstream artifact links used by this project:

• W2V-BERT base weights: https://huggingface.co/facebook/w2v-bert-2.0/resolve/main/model.safetensors?download=true
• Speaker checkpoint: https://huggingface.co/zl389/w2v-bert-2.0_SV/resolve/main/model_lmft_0.14.pth?download=true

GitHub repositories cannot be attached to the Hugging Face base_model field in the same way as Hub models, so the implementation source is listed explicitly above and in artifact_manifest.json.

Runtime contract

This package is a two-flatbuffer pipeline rather than a single direct waveform-to-embedding flatbuffer.

• Sample rate: 16000 Hz
• Stage 1 input: waveform [1, 64000], float32
• Stage 1 output: input_features [1, 199, 160], float32
• Stage 1 preprocessing: SeamlessM4T/Kaldi fbank feature extraction equivalent to the W2V-BERT feature extractor configuration from facebook/w2v-bert-2.0
• Stage 2 input: input_features [1, 199, 160], float32
• Stage 2 output: speaker embedding [1, 256], float32
• Stage 2 quantization mode: dynamic-range
• Recommended postprocessing: L2-normalize each chunk embedding, average chunk embeddings for longer audio, then L2-normalize the pooled vector.

Limitations

• This is an experimental converted artifact, not an upstream official release.
• It is packaged as two TFLite files; callers must run stage 1 before stage 2.
• Quantization can shift speaker similarity scores. Recalibrate thresholds for your deployment.
• The upstream speaker assets are marked in this project as CC BY-NC-SA 4.0 / research/testing unless legal review approves broader use.