w2v-bert-2.0-kinyarwanda-onnx

w2v-bert-2.0-kinyarwanda-onnx

ONNX export of badrex/w2v-bert-2.0-kinyarwanda-asr for use with onnx-asr.

This is the 1st-place solution of the "Kinyarwanda Automatic Speech Recognition — Track A" Kaggle competition, published by Badr M. Abdullah / the ASRwanda team and Digital Umuganda. See the writeup: ASRwanda: 1st place solution for Kinyarwanda ASR Track A.

Kinyarwanda has roughly 15 million speakers. This is the first specialized Kinyarwanda model in the OpenVoiceOS onnx-asr collection.

License and attribution

• License: CC-BY-4.0 (as declared on the source model card), inherited verbatim.
• Base model: facebook/w2v-bert-2.0 (Meta).
• Fine-tuned by Badr M. Abdullah and Digital Umuganda (2025), trained on ~1000 hours of transcribed Kinyarwanda speech from the Kinyarwanda ASR hackathon Track B dataset (Gates Foundation funded, sourced via Digital Umuganda).
• Attribution required per CC-BY-4.0: credit Badr M. Abdullah / the ASRwanda team.

How this was exported

The source model is a Wav2Vec2BertForCTC (Wav2Vec2-BERT 2.0 conformer encoder + CTC head, add_adapter: false, i.e. no additional temporal downsampling beyond the feature frontend).

Its feature frontend is transformers.SeamlessM4TFeatureExtractor: Kaldi-style log-mel filterbanks (Povey window) with per-utterance mean/variance normalization and 2-frame stacking into 160-dim vectors. This is exactly what onnx-asr's w2vbert NumPy preprocessor (W2vBertPreprocessorNumpy) reproduces — verified elsewhere in this collection to ~6e-05 max abs error against the reference feature extractor. So the encoder + CTC head was exported as a single ONNX graph taking the already-stacked 160-dim features (not raw waveform), matching the existing espnet-ctc model contract:

features (B, T, 160) f32, features_lens (B,) i64
  -> logprobs (B, T', V) f32 (log_softmax), logprobs_lens (B,) i64

Loads with:

import onnx_asr
model = onnx_asr.load_model("espnet-ctc", "OpenVoiceOS/w2v-bert-2.0-kinyarwanda-onnx")

(on the feat/espnet-ebranchformer-ctc integration branch of the OpenVoiceOS onnx-asr fork until it lands on dev.)

Subsampling factor: 1 — measured empirically (output frame count equals input feature frame count for several lengths), consistent with add_adapter: false in the source config: the conformer encoder does no extra temporal downsampling beyond the preprocessor's own 2x frame-stacking, so the CTC output rate is one frame per stacked (20ms) feature frame.

Files

• model.onnx (+ model.onnx_data) — fp32 graph, ~2.2GB weights.
• model.int8.onnx — dynamic-range int8 quantized weights.
• vocab.txt — 32-token char vocabulary (kaldi-style token id lines), blank kept at its trained index (<blk> = index 29, the original [PAD] token id).
• config.json — {"model_type": "espnet-ctc", "subsampling_factor": 1}.

Notes

• Not a substitute for anything — this is the actual, HF-published 1st-place competition checkpoint, exported unmodified to ONNX.

Verification

Verified against 3 real Kinyarwanda clips from KYAGABA/kinyarwanda_cleaned_testset_verified (FLEURS has no Kinyarwanda config; Common Voice rw config was not resolvable via the datasets library at test time, so this verified test set was used instead).

• ONNX fp32 (onnx_asr.load_model("espnet-ctc", ...)) output is identical to the native transformers.Wav2Vec2BertForCTC greedy-CTC decode on all 3 clips (byte-for-byte on 2/3; one clip differs only by a collapsed double space) — confirms the ONNX export and the w2vbert preprocessor reproduce the reference model exactly.
• int8 (dynamic-range quantized) shows small, expected degradation vs fp32 on the same clips.

RTF (real-time factor, CPU, OMP_NUM_THREADS=6, on a shared/contended host — treat as upper bounds, not clean single-tenant numbers):

ONNX is consistently faster than native transformers under the same contention, and int8 is faster again than fp32.