theaviv/mevine-0.1

Mevine 0.1

A 107M-parameter Conformer–RNN-T speech recognition model, trained from scratch — no distillation, no pretrained encoder, no fine-tuning of an existing checkpoint.

Open weights, non-commercial research use only. See LICENSE — the restriction is inherited from the training corpus, not a preference. A commercially licensable variant trained without the restricted data is planned.

Results

Open ASR Leaderboard methodology: Whisper EnglishTextNormalizer, corpus-level WER (total edits ÷ total reference words).

dataset	WER	utterances	reference words
librispeech_clean	0.0544	2,620	53,029
librispeech_other	0.1384	2,939	52,884
voxpopuli	0.1167	1,842	44,079
gigaspeech	0.2221	19,898	403,727
earnings22	0.2019	2,731	48,919
spgispeech	0.0743	39,341	969,328
average	0.1346

Throughput

RTFx is audio-seconds decoded per wall-second. Greedy decoding throughout.

GPU	batch	decode	eval set	RTFx
H100 80GB HBM3 (SXM)	16	greedy	gigaspeech	668×
RTX PRO 6000 Blackwell	64	greedy	librispeech_clean	537×
H100 80GB HBM3 (SXM)	16	greedy	spgispeech	462×
RTX PRO 6000 Blackwell	16	greedy	librispeech_clean	394×
H100 PCIe	128	greedy	librispeech_clean	371×
RTX A6000	16	greedy	librispeech_clean	162×

Usage

import torch
from transformers import AutoModel

model_path = "theaviv/mevine-0.1"
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.bfloat16 if device == "cuda" else torch.float32

model = AutoModel.from_pretrained(
    model_path, trust_remote_code=True, dtype=torch_dtype
).to(device).eval()

print(model.transcribe("audio.wav"))

trust_remote_code=True is required: Mevine's architecture is not part of transformers, so the two modules that define it (configuration_mevine.py, modeling_mevine.py) ship in this repo and are loaded from it. Read them — they are the entire model, ~600 lines with no dependencies beyond the list below.

transcribe accepts a file path in any format torchaudio reads, or a 1-D 16 kHz mono float32 tensor. Audio of any length works: clips longer than 30 s are windowed automatically, with cuts nudged into pauses so words are not sliced in half. Pass chunk_s=0 to force a single window.

For batch or frame-level work, call the model directly — model(mel, mel_len) returns token ids, and model.feature_extractor / model.load_audio / model.tokenizer are the pieces transcribe composes.

Requires transformers>=5.0, torch, torchaudio, tokenizers, soundfile, safetensors. Nothing else.

Model

architecture	Conformer encoder (16 layers, d=512, 8 heads) + LSTM prediction net + joint
parameters	107.15M
tokenizer	1024-token byte-level BPE
audio	16 kHz mono; 80-mel log-mel, 25 ms window / 10 ms hop, per-utterance CMVN
decoding	RNN-T greedy; no beam, no external language model, no shallow fusion
precision	bf16

The encoder is offline and bidirectional. Streaming mode is not available as part of open weights release.

Training data

source	hours	licence	dataset
LoquaciousSet (medium+large)	4,883 h	CC-BY-4.0	speechbrain/LoquaciousSet
Earnings-22	119 h	CC-BY-SA-4.0 (share-alike)	sanchit-gandhi/earnings22_split
SPGISpeech	184 h	Kensho research/internal-use, non-commercial	kensho/spgispeech
total	5,167 h		2,058,381 utterances

Read the licence column. SPGISpeech is distributed under a Kensho research/internal-use agreement that restricts redistribution and prohibits commercial use, and Earnings-22 is share-alike. Those terms flow through to these weights. This is disclosed rather than buried: it is why the model is non-commercial, and it is stated in full in LICENSE §3.

GigaSpeech is not in the training corpus — it is gated and contributed 0 hours — which is visible in the results as the weakest column relative to the others.

Limitations

• English only. No other language was trained or evaluated.
• No streaming. The encoder attends bidirectionally over the whole utterance.
• Uneven across domains. Read and prepared speech (LibriSpeech, VoxPopuli, SPGI) are much stronger than spontaneous, noisy or conversational audio (GigaSpeech, Earnings-22). The gap tracks what the corpus contains, not what the architecture can do.
• No punctuation or casing. Output is lowercase, unpunctuated text.
• No speaker diarisation, no timestamps.
• Not evaluated for fairness across accents, dialects or demographics. Do not deploy where that matters without measuring it yourself.
• ASR output contains errors. Do not use for medical, legal, safety-critical or surveillance purposes.

Training

hardware	1× H100 80GB
wall time	15.7 hours
steps	200,000
effective throughput	3.56 steps/s
optimiser	AdamW, Noam schedule (peak 3e-4, 2000 warmup)
loss	0.7 × RNN-T + 0.3 × CTC auxiliary (the CTC head is discarded at inference)
stability	0 OOM recoveries; 74 dropped items of 2.06M

The dev-WER curve was still descending when the compute budget ran out — 200k steps was a budget, not convergence.

Commercial use

These weights are research-only for the corpus reason above, not because the technology is unavailable. A commercially licensable variant — same architecture, same training system, retrained without the restricted data — is in progress.

What is actually on offer is the system rather than this checkpoint: the corpus pipeline, the evaluation harness and the training loop were built to be pointed at a specific domain and re-run, which is what moves WER on audio that a general model handles badly. Mevine 0.1 is the evidence that the loop works end to end, from random initialisation.

If you have an audio workload you would want a model tuned for, open a discussion on this repo.

Citation

@misc{{mevine2026,
  title  = {{Mevine 0.1: a from-scratch 107M Conformer-RNN-T for English ASR}},
  author = {{Dozorets, Aviv}},
  year   = {{2026}},
  note   = {{Open weights, non-commercial research licence}},
  url    = {{https://huggingface.co/theaviv/mevine-0.1}}
}}

Acknowledgements

Trained on LoquaciousSet (SpeechBrain), Earnings-22 (Rev.com), and SPGISpeech (Kensho). Evaluated with the Open ASR Leaderboard methodology.