---
title: 'ACL 60/60 Corpus: FeruzaSpeech for Uzbek ASR'
url: https://www.emergentmind.com/topics/acl-60-60-corpus
type: topic
---

# ACL 60/60 Corpus: FeruzaSpeech for Uzbek ASR

The designation **ACL 60/60 Corpus** is not presented as a universally standardized dataset name in the cited literature. In the ACL / speech-NLP sense, the closest documented referent is **FeruzaSpeech**, explicitly introduced as “a 60 hour Uzbek read speech corpus” and described as essentially a “60/60-style” resource: a carefully documented, approximately 60-hour corpus with well-specified splits, scripts, and baseline ASR results [2410.00035]. FeruzaSpeech consists of 59.61 hours of high-quality Uzbek read speech from a single native female speaker from Tashkent, preserves punctuation, casing, and multi-sentence context, and provides aligned Latin and Cyrillic transcripts. By contrast, in medical AI and biomechanics, “ACL” refers to the anterior cruciate ligament; resources such as ACL27 and simulation-based ACL injury studies are unrelated to this speech-corpus usage and do not define a dataset by the name “ACL 60/60 Corpus” [2502.07431] [2202.13749].

## 1. Corpus identity and intended function

FeruzaSpeech is a read-speech corpus of the Uzbek language with transcripts in both **Latin** and **Cyrillic** alphabets. It is built from recordings by a **single native female speaker** from **Tashkent, Uzbekistan**, representing standard Uzbek in a controlled read-speech setting. The textual material comes from two sources: excerpts from the classic Uzbek novel *Choliqushi* and **BBC Uzbek News** articles. Its documented duration is

$$
\text{Book} = 21.57\,\text{h},\quad \text{BBC Uzbek} = 38.04\,\text{h},\quad \text{Total} = 59.61\,\text{h}.
$$

Throughout the paper, this is rounded to “60 hours” [2410.00035].

The corpus is positioned as a resource for Uzbek **ASR** and **TTS** in a low-resource setting. The stated goals are to fill a resource gap for Uzbek speech technology, provide the **first Uzbek corpus with both Latin and Cyrillic transcriptions**, and indeed the first with **Cyrillic transcriptions at all**, retain **punctuation, casing, and multi-sentence context** rather than normalized text, and demonstrate that adding this 60-hour corpus improves ASR performance on multiple Uzbek datasets. In the ACL-style research sense, its defining characteristics are not merely duration but documentation, explicit train/dev/test splits, and reproducible baselines [2410.00035].

## 2. Textual design, annotation regime, and dual-script representation

A distinctive property of FeruzaSpeech is that it preserves **“natural” text**. Punctuation is retained, casing is preserved rather than lowercased, and the segments are deliberately longer than conventional short-form ASR cuts. Each audio segment contains **1–2 full sentences**. The reported average segment length is **16.39 seconds**, with a minimum of **3.78 s** and a maximum of **50.69 s**. The paper notes that these segments are consistently longer than those of the Uzbek Speech Corpus, whose segments are mainly **2–3 seconds** [2410.00035].

This annotation policy is explicitly contrasted with more aggressively normalized corpora. FeruzaSpeech preserves punctuation and casing in the transcript itself rather than treating them as pre- or post-processing artifacts. The authors argue that **modern deep models can handle unnormalized text**, and that this reduces dependence on explicit normalization and inverse text normalization stages. A plausible implication is that the corpus is intended not only for lexical recognition but also for experiments in joint orthographic modeling, punctuation-aware decoding, and context-sensitive generation [2410.00035].

The corpus also implements a dual-script transcription design. The original texts were in **Cyrillic**, read by the speaker, then converted to **Latin** using online converters, specifically **Lexilogos** and **uzlatin.com**, followed by **manual correction** of conversion errors. The final dataset provides **both Latin and Cyrillic transcripts aligned at the segment level**, so each audio file is paired with orthographically equivalent content in both scripts. No phonetic labels, prosodic annotations, POS tags, syntactic parses, or other linguistic layers are reported; the core annotation is the **word-level orthographic transcript in two scripts with punctuation and casing** [2410.00035].

## 3. Acoustic format, segmentation policy, and reproducible splits

The recordings are distributed as individual `.wav` files in a controlled acoustic format: **single-channel (mono)**, **16-bit**, **16 kHz**. The paper describes the environment as high quality and explicitly states that there is **no background noise**. Although no microphone or hardware specification is given, the combination of a single careful reader, quiet recording conditions, and strong ASR results functions as the practical quality-control regime [2410.00035].

FeruzaSpeech is pre-split into **Train**, **Dev**, and **Test** sets. The durations are **52.09 h** for Train, **2.93 h** for Dev, and **4.08 h** for Test. The **Dev** and **Test** sets contain only **BBC Uzbek News**, whereas **Train** contains both **BBC** and **Choliqushi** material. This creates a deliberately asymmetric domain design: literary material is available during training, while evaluation is restricted to the news domain. The setup therefore measures both in-domain recognition of news and any regularization benefit contributed by the literary portion of the corpus [2410.00035].

For experimental comparison, the paper also uses **Common Voice 16.1 (Uzbek)** with **54.88 h** of training data and the **Uzbek Speech Corpus (USC)** with **90.70 h** of training data; the combined **CV + FS + USC** training set reaches **197.68 h**. The dataset is described as **freely available for academic research purposes** and hosted on **HuggingFace Datasets** under `k2speech/FeruzaSpeech` at `https://huggingface.co/datasets/k2speech/FeruzaSpeech`. Reproducible training configurations are linked through public **Icefall** recipes, including the Common Voice RNN-T Conformer recipe and the LibriSpeech Zipformer recipe [2410.00035].

## 4. Baseline ASR methodology and reported performance

The reported ASR experiments are conducted in **Next-Gen Kaldi / Icefall**. Two model families are used. The first is a **Stateless RNN-T Conformer** based on the `pruned_transducer_stateless7` recipe, using a Conformer encoder and evaluated with both **greedy search** and **modified beam search**. The second is **Zipformer**; for Uzbek, the paper uses **small zipformer** parameters because the default Common Voice recipe did not converge with full Zipformer settings. All reported models are trained for **60 epochs** [2410.00035].

Evaluation is performed under two scoring regimes. **C&P** scores with **Casing and Punctuation**, i.e. exact orthographic match including punctuation and letter case. **UNP** denotes **Uppercase, No Punctuation**, meaning that text is normalized at scoring time. The evaluation metric is **Word Error Rate (WER)**:

$$
\text{WER} = \frac{S + D + I}{N} \times 100\%,
$$

where \(S\) is substitutions, \(D\) deletions, \(I\) insertions, and \(N\) the number of reference words [2410.00035].

The central empirical result is that **adding FeruzaSpeech improves ASR performance on multiple Uzbek test sets**, despite its single-speaker design. For the **Stateless RNN-T Conformer** with **modified beam search**, the **C&P** WER on **cv-test** improves from **31.98** with **CV** training alone to **30.47** with **CV+FS**; on **usc-test** it improves from **51.61** to **48.60**. Under **UNP** scoring, the corresponding **cv-test** WER improves from **20.16** to **18.33**, and **usc-test** from **34.03** to **29.67**. When **CV+FS+USC** are combined, the best reported WERs are **27.81** on **cv-test** and **9.56** on **fs-test** under **C&P**, and **11.17** on **cv-test**, **4.05** on **fs-test**, and **11.67** on **usc-test** under **UNP**. The paper further states that, relative to the original USC paper’s best **17.4%** WER on **usc-test**, the combined model’s **11.67%** constitutes an **absolute reduction of 5.73 points**, or about a **33% relative improvement** [2410.00035].

The **Zipformer** experiments show the same direction of effect, though with smaller gains on cross-corpus generalization. Under **C&P** scoring with **modified beam search**, **cv-test** improves from **33.96** with **CV** alone to **33.15** with **CV+FS**, **fs-test** from **32.41** to **10.75**, and **usc-test** from **54.07** to **53.08**. Under **UNP**, **cv-test** improves from **21.92** to **21.35**, **fs-test** from **20.21** to **5.10**, and **usc-test** from **37.44** to **35.94**. The paper summarizes these trends by stating that adding FeruzaSpeech to Common Voice improves WER by **1.49–2.12** points on **cv-test** and **3.01–4.58** points on **usc-test** for the Stateless RNN-T Conformer, and by **0.57–1.10** points on **cv-test** and **0.54–1.50** points on **usc-test** for Zipformer [2410.00035].

## 5. Relation to existing Uzbek corpora and the “60/60-style” characterization

FeruzaSpeech is explicitly positioned as complementary to two other Uzbek resources. **Common Voice 16.1 (Uzbek)** is described as having around **265 hours** total across train/dev/test, more than **2,000 speakers**, crowd-sourced and more variable recording conditions, mostly in **Latin** script, and generally shorter segments. **USC** is described as around **105 hours**, with **958 speakers**, normalized text **without casing and punctuation**, and typical segments of **2–3 seconds** [2410.00035].

Within this landscape, FeruzaSpeech contributes several properties absent or weakly represented elsewhere: **Cyrillic transcriptions**, **natural punctuation and casing**, **longer contextual segments**, and **high acoustic quality** with minimal noise. At the same time, it lacks **speaker diversity** and **acoustic/environmental variability**. The paper therefore presents it less as a replacement for multi-speaker corpora than as a high-quality complementary resource that can be combined with them. The best-performing systems are obtained precisely through **data combination**, especially **CV + FS + USC**, rather than from FeruzaSpeech as a stand-alone ASR training set [2410.00035].

This framing explains the “60/60-style” label. In the paper’s own positioning, FeruzaSpeech functions as a **60-hour, high-quality, single-speaker, read ASR corpus with punctuation, casing, and context**, published with the documentary and experimental apparatus expected of ACL-style speech research. It is therefore a natural candidate when the phrase “ACL 60/60 corpus” is used informally to mean a roughly **60-hour** corpus that is **ACL-ready** in the sense of documentation, public availability for research, explicit splits, and baseline ASR experiments [2410.00035].

## 6. Terminological ambiguity, limitations, and future directions

The phrase **ACL 60/60 Corpus** is complicated by acronym ambiguity. In speech and NLP, “ACL” commonly evokes the Association for Computational Linguistics and ACL-style publication and benchmarking practices. In medical vision, however, “ACL” denotes the **Anterior Cruciate Ligament**. The paper **“ArthroPhase: A Novel Dataset and Method for Phase Recognition in Arthroscopic Video”** introduces **ACL27**, a dataset of **27** arthroscopic ACL reconstruction videos with five surgical phases, but explicitly states that it does **not** mention an “ACL 60/60 Corpus” by name. Similarly, the thesis **“Predictive simulation of single-leg landing scenarios for ACL injury risk factors evaluation”** provides a simulation pipeline and structured ACL risk metrics for biomechanics, not a speech corpus or a dataset bearing that designation [2502.07431] [2202.13749]. This suggests that the expression is better understood as a contextual shorthand than as a formally registered cross-domain corpus name.

FeruzaSpeech itself has explicitly stated limitations. It is a **single-speaker** corpus from **one female speaker**, contains **read speech only**, includes **no background noise**, and uses relatively **long segments** with an average duration of **16.39 seconds**, which some ASR systems may need to segment further. Although both **Latin** and **Cyrillic** transcripts are provided, the paper’s experiments are run only on **Latin** transcripts, so Cyrillic usage remains unexplored experimentally. The paper identifies several future directions: providing the data at **higher sampling rates and bit depth** for TTS, expanding the corpus with additional recordings from the **same speaker**, conducting explicit **TTS experiments**, and improving tools for **accurate Latin–Cyrillic conversion**, especially for characters such as the **soft sign (ь)** [2410.00035].

Taken together, the available literature supports a narrow and technically specific reading of the term. In that reading, an **ACL 60/60 Corpus** is not a canonical named benchmark but a description that fits FeruzaSpeech unusually well: a roughly **60-hour**, **freely usable for academic research**, **well-specified** Uzbek read-speech corpus with dual scripts, punctuation, casing, long-context segmentation, and published ASR baselines.

Source: https://www.emergentmind.com/topics/acl-60-60-corpus