OLMoASR-Pool: Open Data for Robust ASR

• The paper demonstrates that rigorous filtering of large-scale weakly supervised data reduces WER and enhances zero-shot ASR performance.
• OLMoASR-Pool is defined as a diverse web-crawled corpus of speech–text pairs, which, when curated into OLMoASR-Mix, provides high-quality training data.
• Ablation studies confirm that sequential filtering stages—from segmentation to language alignment—significantly improve ASR accuracy across model scales.

Searching arXiv for the cited OLMoASR paper to ground the article in the latest available record. OLMoASR-Pool is a large-scale collection of weakly supervised English speech–text pairs introduced in "OLMoASR: Open Models and Data for Training Robust Speech Recognition Models" (Ngo et al., 28 Aug 2025). It comprises approximately 3 million hours of English audio paired with 17 million transcripts and serves as the starting point for studying how data scale and curation affect zero-shot automatic speech recognition (ASR). From this pool, the paper derives a filtered corpus, OLMoASR-Mix, containing approximately 1 million hours of high-quality audio–text pairs, and uses that corpus to train the OLMoASR model family, a suite of Whisper-style encoder-decoder ASR models spanning 39 M to 1.55 B parameters. The central empirical claim is that careful filtering of large, weakly labeled web-scale data materially improves robustness and yields average zero-shot performance that is comparable to OpenAI’s Whisper on both short-form and long-form benchmarks (Ngo et al., 28 Aug 2025).

1. Corpus definition and source composition

OLMoASR-Pool is defined as a heterogeneous “web crawl” of publicly accessible speech–text pairs. The examples given include podcast auto-generated captions, publicly uploaded transcripts to social-media or news sites, audiobooks, and lecture recordings. All of these sources were harvested under weak supervision, with no manual re-alignment. In that sense, OLMoASR-Pool is not a hand-curated transcription corpus; it is a large, open pool of weakly labeled speech designed for subsequent filtration and analysis (Ngo et al., 28 Aug 2025).

The stated purpose of the pool is to serve as an open, large-scale source of weakly labeled speech for studying how data scale and curation affect zero-shot ASR. This framing is important because the paper does not present the raw pool as a final training set. Instead, it treats the pool as a broad acquisition layer from which more reliable training data can be distilled.

A common misconception is that the value of such a resource lies only in its absolute scale. The paper’s design instead places equal emphasis on scale and curation. This suggests that OLMoASR-Pool should be understood less as a benchmark-ready dataset than as a substrate for investigating filtering policies, contamination control, and robustness-oriented training.

2. Curation from OLMoASR-Pool to OLMoASR-Mix

The conversion of OLMoASR-Pool into OLMoASR-Mix proceeds through a sequence of filtering stages. The paper begins with 3 M hours, applies segmentation, enforces audio–text language agreement, then uses text heuristics, and finally performs fuzzy de-duplication and de-contamination. The result is a new dataset containing approximately 1 million hours of high-quality audio–transcript pairs, called OLMoASR-Mix (Ngo et al., 28 Aug 2025).

Stage	Hours remaining	Key operation
Start	3 M	Raw OLMoASR-Pool
Segmentation	2.4 M	Uniformly segment long audio files into utterances
Audio–text language alignment	2.0 M	VoxLingua107 for audio, pycld2 for text, retain only “English”/“English” pairs
Repeating-lines filter alone	1.21 M	Drop samples whose transcripts contain identical consecutive lines
Manual–machine WER filter	0.91 M	Apply document-level and segment-level transcript disagreement thresholds
Final OLMoASR-Mix	≈1 M	Combine repeating-lines removal, WER filtering, de-duplication, and de-contamination

The segmentation stage uniformly segments long audio files into utterances and retains 2.4 M hours, or 80 % of the original pool. Audio–text language alignment then predicts the top-1 spoken language from audio with VoxLingua107 and the top-1 text language with pycld2, retaining only pairs where both predict English; this produces 2.0 M hours, approximately 83 % of 2.4 M.

The paper’s final policy for OLMoASR-Mix combines two main text-heuristic filters. The first removes samples whose transcripts contain identical consecutive lines. The second generates a “noisy” ASR transcript for each segment using an open-source speech recognizer and computes transcript disagreement at two granularities. At the document level,

$$\mathrm{WER}_{doc} = \frac{\mathrm{edit\text{-}distance(manual, machine)}}{|manual|},$$

and samples are dropped if $\mathrm{WER}_{doc} < 0.5$. At the segment level, the analogous $\mathrm{WER}_{seg}$ is computed and segments are dropped if $\mathrm{WER}_{seg} < 0.7$.

After these heuristic stages, the paper applies fuzzy de-duplication and de-contamination. It uses MinHash on 5-grams of tokens, with 112 hashes in 14 buckets, to remove transcript pairs above Jaccard approximately 75 %, affecting approximately 3 % of 17 M transcripts. It also performs n-gram lookup to remove any transcript containing 10-grams that appear in evaluation sets, reporting approximately 0.002 % contamination. The net effect is a final OLMoASR-Mix of approximately 1 million hours of high-quality audio–text pairs.

3. Filtering ablations and data-quality effects

The paper reports ablations that quantify how quality filtering changes downstream ASR performance. On a 74 M-parameter baseline model, the short-form average word error rate (WER) over 14 tasks is 37.2 % with no quality filtering. After uppercase/lowercase removal, WER becomes 32.4 %, with 32 % of the data removed. After repeating-lines removal, WER becomes 22.7 %, with 39.9 % of the data removed. After the manual–machine filter, WER becomes 20.7 %, with 54.8 % removed (Ngo et al., 28 Aug 2025).

These numbers establish that the paper’s argument is not simply that more weakly supervised data improves ASR. Rather, the reported WER trajectory indicates that filtering policy has a large effect on training utility. The repeating-lines heuristic alone retains 1.21 M hours, or 60 % of the 2.0 M hours remaining after language alignment, and improves short-form WER from 37.2 % to 22.7 % on 14 held-out tasks. The manual–machine WER filter retains 0.91 M hours, or 45 % of 2.0 M hours, and further reduces short-form WER to 20.7 %.

A plausible implication is that OLMoASR-Pool’s importance lies partly in enabling high-recall acquisition before quality selection. The paper’s evidence suggests that web-scale weak supervision can be productive for zero-shot ASR, but only after aggressive curation. This directly counters the view that ASR training quality is primarily limited by model architecture rather than by the interaction between scale, noise, and filtering.

4. OLMoASR models trained on OLMoASR-Mix

OLMoASR-Mix is used to train the OLMoASR-Mix suite of models, which are described as Whisper-style encoder-decoder models with FlashAttention and causal/padding masks. The model sizes are tiny.en at 39 M parameters, base.en at 74 M, small.en at 244 M, medium.en at 769 M, and large.en / large.en-v2 at 1.55 B (Ngo et al., 28 Aug 2025).

The training hyperparameters are the same for all scales. The optimizer is AdamW with $\beta_1 = 0.9$, $\beta_2 = 0.98$, $\epsilon = 1e\!-\!6$, and weight decay $= 0.1$. Training runs for 524,288 updates with batch size $= 512$, warmup $= 1,049$, and max-LR $\mathrm{WER}_{doc} < 0.5$0 with linear decay. Training precision is FP16 with DDP for small models and bfloat16 with FSDP for large models. The objective is standard next-token cross-entropy on the concatenated text tokens.

The paper reports a zero-shot impact from curation: models trained on curated OLMoASR-Mix outperform identical-compute baselines trained only on a non-quality-filtered subset, reducing both short-form and long-form WER by several percentage points across all sizes. This makes OLMoASR-Pool and OLMoASR-Mix jointly significant: the pool enables scale, while the curated mix supplies the higher-quality training signal that the experiments associate with improved zero-shot performance.

5. Comparative zero-shot performance and robustness

Across all model scales, OLMoASR achieves comparable average performance to OpenAI’s Whisper on short-form and long-form speech recognition benchmarks. The paper emphasizes a parameter-matched comparison at the medium scale: OLMoASR-medium.en attains 12.8 % short-form WER and 11.0 % long-form WER, compared with Whisper-medium.en at 12.4 % and 10.5 %, respectively (Ngo et al., 28 Aug 2025).

Model	Short-form average WER	Long-form average WER
tiny.en	OLMoASR 20.5 % vs. Whisper 20.1 %	OLMoASR 15.6 % vs. Whisper 16.6 %
base.en	OLMoASR 16.6 % vs. Whisper 16.9 %	OLMoASR 12.9 % vs. Whisper 13.2 %
small.en	OLMoASR 13.8 % vs. Whisper 13.7 %	OLMoASR 11.5 % vs. Whisper 11.2 %
medium.en	OLMoASR 12.8 % vs. Whisper 12.4 %	OLMoASR 11.0 % vs. Whisper 10.5 %
large.en	OLMoASR 13.0 % vs. Whisper-large-v1 12.2 %	OLMoASR 11.4 % vs. Whisper-large-v1 10.4 %

For the large scale, the paper reports that re-trained large-v2 on 680 K h yields 12.6 % short-form WER, bringing it closer to parity with Whisper-large-v1. On out-of-distribution sets—AMI, CHiME-6, CORAAL, CommonVoice, and VoxPopuli—OLMoASR exhibits positive effective robustness, meaning it outperforms its in-distribution slope, and positive relative robustness compared to supervised LibriSpeech-only models.

These results situate OLMoASR-Pool within a broader robustness narrative. The paper does not claim universal superiority over Whisper at every scale and benchmark. Instead, it argues that an open, carefully curated training pipeline can approach or match Whisper-level average performance while remaining publicly available for research use. That distinction matters for researchers interested in data-centric ASR development rather than only in closed-model replication.

6. Availability, licensing, and research significance

The paper states that OLMoASR-Pool and the OLMoASR-Mix manifest, defined as lists of audio-file IDs, are published on HuggingFace at https://huggingface.co/datasets/allenai/OLMoASR-Pool and https://huggingface.co/datasets/allenai/OLMoASR-Mix. Filtering, training, and evaluation code are available on GitHub at https://github.com/allenai/OLMoASR, and trained model weights are available on HuggingFace at https://huggingface.co/allenai/OLMoASR (Ngo et al., 28 Aug 2025).

The release is for academic research only. The paper further recommends that users evaluate legal, privacy, and fairness implications before any deployment. This is an important qualification: open availability does not imply unrestricted operational use, and the paper explicitly frames downstream deployment as requiring additional scrutiny.

In aggregate, OLMoASR-Pool represents a 3 M-hour, 17 M-transcript weakly supervised corpus whose main scientific role is to enable controlled study of scale, noise, and curation in zero-shot ASR. The derivation of OLMoASR-Mix from that pool, together with the performance of the resulting OLMoASR models, suggests that large-scale web-supervised speech resources can support robust ASR research when accompanied by explicit alignment checks, heuristic transcript filtering, fuzzy de-duplication, and de-contamination.