yehoshua01/waxal-whisper-turbo-lin-r1

Lingala ASR — whisper-large-v3-turbo fine-tune, round 1

Component of the 4th-place solution to the Google WAXAL ASR Challenge (Zindi, phase 2): 892 unseen clips, two African languages, no language metadata, scored 1 - (WER + CER) / 2 on raw text. Private leaderboard 0.771848284.

Code, full method and one-command verification: [yehoshua0/waxal-asr-phase2](https://github.com/yehoshua0/waxal-asr-phase2) The repository reproduces the submitted CSV byte for byte on a laptop in about a minute, and re-decodes every input from the audio on rented GPUs in about four hours.

Role in the system

Witness (voter), and the one the shipped chain actually reads. Decode it under the placeholder language token <|sw|>, which is what it trained under: a language token is a learned decoder state, and <|ln|> selects one this checkpoint never saw.

What it measured

Beam 3 gains +0.003369 over greedy on this checkpoint. As a solo Lingala half it scores 0.752910, i.e. -0.009491 against the shipped chain.

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch, soundfile as sf, torchaudio.functional as AF

proc = WhisperProcessor.from_pretrained("yehoshua01/waxal-whisper-turbo-lin-r1")
model = WhisperForConditionalGeneration.from_pretrained("yehoshua01/waxal-whisper-turbo-lin-r1").eval().cuda()
tok = proc.tokenizer
# the language slot is a LEARNED decoder state -- use the token this checkpoint trained under
forced = [(1, tok.convert_tokens_to_ids("<|sw|>")),
          (2, tok.convert_tokens_to_ids("<|transcribe|>")),
          (3, tok.convert_tokens_to_ids("<|notimestamps|>"))]

wav, sr = sf.read("clip.wav", dtype="float32")
wav = AF.resample(torch.from_numpy(wav), sr, 16_000).numpy()[:30 * 16_000]
f = proc.feature_extractor(wav, sampling_rate=16_000, return_tensors="pt").input_features
out = model.generate(f.to("cuda", model.dtype), forced_decoder_ids=forced, max_new_tokens=220)
print(proc.batch_decode(out, skip_special_tokens=True)[0])

The rest of the system

Sibling checkpoints (primaries, voters and ablations of the same system): `waxal-mms-1b-lin-pl2-spk` · `waxal-sunbird51-sna-pl2-spk` · `waxal-whisper-turbo-lin-r2` · `waxal-qlora-largev3-lin` · `waxal-omni-ctc1b-lin` · `waxal-omni-ctc1b-sna` · `waxal-sunbird51-lin-ft-r2` · `waxal-sunbird51-lin-ft-light` · `waxal-mms-1b-lin-full` · `waxal-mms-1b-lin-fullmeta` · `waxal-ssa-hubert-lin`

Licence and intended use

apache-2.0. Training data is `google/WaxalNLP` (CC-BY-SA-4.0, share-alike), so derivatives carry that too.

These weights are not a general-purpose ASR model. Pseudo-labels were computed on the phase-2 test audio (transductive self-training, permitted for phase-2 training by the host), so the checkpoint is partly adapted to that specific set.