Universal Representations of Speech](https://arxiv.org/abs/2205.12446) Fleurs is the speech version of the FLoRes machine translation benchmark. We use 2009 n-way parallel sentences from the FLoRes dev and devtest publicly available sets, in 102 languages. Training sets have around 10 hours of supervision. Speakers of the train sets are different than speakers from the dev/test sets. Multilingual fine-tuning is used and ”unit error rate” (characters, signs) of all languages is averaged. Languages and results are also grouped into seven geographical areas: The datasets library allows you to load and pre-process your dataset in pure Python, at scale. The dataset can be downloaded and prepared…
SAVRN Model Hub · Datasets by Task
Speech Recognition Datasets
29 open-weight speech recognition datasets in the SAVRN Model Hub, with Shrikant Nayak, The Larry Project and Thant Zin Phyo publishing the most.
29Datasets
20Publishers
7Licenses
SAVRN's Take
The speech recognition datasets here are few but telling. Google's FLEURS, which covers about a hundred languages with the same sentences read aloud, is the standard multilingual benchmark. The rest are mostly Indian-language collections, including call-center speech and university-contributed recordings, which reflect how much open speech work is now aimed at underserved languages.
Most of this category is CC-BY-4.0, so commercial use is allowed with attribution. Speech data carries extra obligations beyond the license: speakers' voices are personal data in many jurisdictions. Check how consent was collected before training on recordings of real people, especially call-center audio.
Most Downloaded
| Dataset | Publisher | License | Monthly downloads |
|---|---|---|---|
| fleurs | cc-by-4.0 | 102.1k | |
| Quranic-Recitation-Data | Muhammad Zohaib Hassan | apache-2.0 | 22.4k |
| bolAIndia | Kapture CX | Not stated | 18k |
| balalaika-sft-rover7-punct | Lab260 | Not stated | 7.2k |
| QuranLab-Arabic-Speech | Mostafa Mahdi | other | 3.9k |
| unnamed-split-ar | Ahmed Wasfy | cc-by-3.0 | 3.6k |
| dialectal-arabic-voices | Moaead | Not stated | 1.7k |
| Vartalaap | Kapture CX | cc-by-nc-4.0 | 1.7k |
| DNU-ViSpeech | Trang Lê | cc-by-nc-4.0 | 1.4k |
| Dhravani | Shreesha | cc-by-4.0 | 1.2k |
Licenses
| License | Datasets | Commercial use |
|---|---|---|
| cc-by-4.0 | 12 | Yes |
| other | 5 | Read the license |
| not stated | 5 | Not stated |
| cc-by-nc-4.0 | 3 | Not without separate permission |
| cc-by-3.0 | 2 | Read the license |
| cdla-sharing-1.0 | 1 | Read the license |
| apache-2.0 | 1 | Yes |
Who Publishes Them
| Publisher | Datasets |
|---|---|
| Shrikant Nayak | 6 |
| The Larry Project | 2 |
| Thant Zin Phyo | 2 |
| Ahmed Wasfy | 2 |
| Kapture CX | 2 |
| Muhammad Zohaib Hassan | 1 |
All 29 Datasets
India does not speak in a single voice. It speaks through a vast and vibrant mosaic of languages, regional accents, dialects, and everyday expressions—across cities, towns, and rural communities. bolAIndia is a large-scale, expanding collection of Indian speech built to help speech AI hear that richness. It brings together conversational, read, educational, and broadcast audio, including code-switching and regional varieties, in a consistent, filterable format. Our ambition is to help researchers build speech technologies that serve more of India: across linguistic, geographic, and social boundaries. Breadth is a continuing goal, not a claim of complete or demographically balanced coverage.…
Dual-channel synthetic Indian customer-support calls for training full-duplex speech-to-speech models. 81,152 calls · 1,880.2 hours · 2201 shards (last updated 2026-10-05 09:06 IST) Each row's audio is a stereo FLAC at 24000 Hz: Hindi is written in Devanagari, English in Latin, in the same string — "आपका order अभी out for delivery है". Romanized Hindi is rejected by a QA gate. Everything the TTS spoke is spelled out: no digits (₹4,500 → चार हज़ार पाँच सौ रुपये), abbreviations letter-by-letter in Devanagari (OTP → ओ टी पी), phone numbers in two-digit batches (अट्ठानवे चालीस पैंतालीस बावन तिरासी). Backchannels are deliberately Latin (ji, hmm ji, haan ji). - callid, audio, durations…
Overview Quranic Recitation Dataset (Word-by-Word Sync) is a highly optimized, production-ready dataset containing high-quality audio recitations of the Holy Quran synchronized at the word-by-word level. This dataset features 135 world-renowned reciters, with every Surah (114 chapters) mapped precisely to millisecond-accurate word timestamps. It is designed for modern Islamic mobile and web applications — served via a Cloudflare Edge CDN with native
Verified public-resume snapshot: 2026-10-10 10:11:08 UTC; 336,830 examples from 420 / 7,797 source shards. All 420 have independent source/output verification; 1,061 remote files were checked. The full run continues on physical GPU 0/1 with a 40 GB node RAM limit. See provenance/audits/publicresume420.json. On 2026-10-10, the user requested making this repository public and continuing uploads after lab260 reached its private storage limit. The last snapshot before this change covers 335,758 uploaded examples from 418 / 7,797 source shards, independently verified against sources and 1,057 remote files. Another 1,072 examples in two audited local shards were uploaded from their preserved…
Broad Arabic speech for next-generation ASR. Access is restricted to NPL1.2 and the internal Quran Lab team. - Broad Arabic speech in segments of 2 to 20 seconds, cut at natural pauses (silences of 1.5 s or longer are removed), sized for CTC/transducer training and for LLM transcription. - Audio is 48 kHz Opus, exactly as the source streamed it (never re-encoded). Config tts holds studio Arabic podcasts and professional narration. Each recording is graded for TTS use on four 9 s windows inside its speech, and the grade is in metadata/recordings.parquet: Recordings with music under the speech (music score above 0.15) or under 75% speech also fail. Earlier broad recordings from the same…
The data/ar/ partition of espnet/yodas3, cut from long recordings (median ~5 minutes) into clips of 5-45 seconds, each with its own transcript. Ready to batch for ASR training without further segmentation. Audio is Opus 48 kbps mono at 16 kHz in Ogg. Upstream timestamps are YouTube caption cues, not utterance segments. Measured on a shard of data/ar: 71.8% of consecutive cues strictly overlap (median gap -2.36 s): start + duration is how long a cue stays on screen, not when its speech ends. Only 9.7% of cue boundaries have a real silence gap >= 0.3 s, so cutting on caption boundaries puts most cuts inside continuous speech. Snapping caption boundaries to the quietest nearby window does not…
An expanding collection of Arabic audio from YouTube, SoundCloud, and other sources. Currently labelled Palestinian Arabic (ps). 53,194 recordings · approximately 9,573.3 hours · 501.14 GB Audio retains its original format and sample rate. Recordings have not yet been segmented or individually checked for dialect. New audio is added in ~5 GB Parquet files. The final incomplete file is held locally until it fills.
DNU-ViSpeech is a Vietnamese scripted/read-speech corpus for automatic speech recognition (ASR). This release contains 31,070 utterances from 26 anonymized speakers, totaling 75.570 hours. Audio is WAV, 16 kHz, mono, uncompressed PCM 16-bit. The primary protocol is DNU-ASR-v1.1. Its evaluation split combines unseen-speaker and recording-configuration shift; those two factors cannot be isolated with the current data. The corpus supports research on Vietnamese ASR adaptation, especially whether compact models can approach larger zero-shot systems after domain fine-tuning on read speech. The repository follows the Hugging Face AudioFolder layout. Each split has its own metadata.csv, and…
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication. - ⌨ Keyboard shortcuts for efficiency 1. Create a transcript CSV file with your content: 2. Start the Flask application: 3. Access the interface: 1. Authentication 2. Session Setup - Click "Start Session" 3. Recording - Use on-screen controls or keyboard shortcuts: - R: Start recording / Stop recording - Space: Play recording - Enter: Save…
Append-only teacher-output cache used to distill a ternary Hindi ASR student. It stores, per audio clip, ONLY: - rowid (a deterministic source-shard/position pointer), - top-k (k=64) CTC logits (vocab id + log-prob per kept entry, blank always kept), - selected encoder hidden states (last-3 blocks: layers 14 / 15 / 16), fp16. No audio and no ground-truth transcripts are stored or redistributed. - Source corpora (audio & transcripts NOT redistributed here; used only to compute teacher outputs): - ai4bharat/Kathbath (hindi) — CC-BY-4.0. - ai4bharat/Shrutilipi (hindi) — CC-BY-4.0. CC-BY-4.0. Redistributing the derived teacher outputs (logits + hidden states + ids) is not redistribution of the…
This repository is configured as public with manual access approval. No audio, transcripts, or training manifests have been uploaded yet. The original transfer-staging metadata is retained below as history. Since 2026-10-08 this repository also preserves encrypted saved fullstate checkpoints from the actual 5000h development training and its continuations. Source: ReazonSpeech, provided by Reazon Human Interaction Laboratory / Reazon Holdings, Inc. Upstream specifies use solely within Article 30-4 of the Japanese Copyright Act; review the upstream dataset card and comply with applicable law. The license is not altered by this access gate. No additional license restrictions are asserted…
Upload in progress. This repository is not yet complete and should not be used or cited until this notice is removed. Opus-compressed audio with rich captions and curation subset labels, in the archive layout the ESPnet speechlm dataloader reads directly. This is the compressed, directly trainable counterpart to which ships HuggingFace-style Parquet that the training code cannot load. each utterance carrying its rich caption as a plain string and its membership in the curation subsets (speechund, speechgen, soundund, soundgen, musicund, musicgen, plus the SFT-sized variants). An utterance belongs to zero, one or two of the base subsets. A full dataset card, per-dataset statistics and usage…
Audio from Uzbek-language YouTube podcasts / audiobooks / fairy tales, in original quality (no re-encoding), together with YouTube subtitles: manual tracks and original-language auto-captions with word-level timings. audio/ / /.opus|.m4a subtitles/ / /.json — {"tracks":[{"lang","kind":"manual|auto","segments":[{"start","end","text","words":[{"start","end","w"}]}]}]} metadata/part-.jsonl — one row per video (title, channel, duration, paths, subtitle tracks) Rights to the content belong to the original authors/channels; intended for research use.
Built from the URL manifest at thelarryproject/scotus-manifest. One row per oral argument session of the U.S. Supreme Court, October Terms 2010–2025, materializing the official audio recording and transcript published at supremecourt.gov. Contains ~1,215 hours of audio un-pruned. Some amount of work has been performed to format the PDF text into a transcript consisting of a series of speaker turns. - docket: docket number; rearguments and multi-session arguments carry suffixes (rearg, q1, q2) - term: October Term the argument belongs to - dateargued: ISO date of the argument session - pageurl: url for transcript and mp3 for a term and docket, in the format…
Newly acquired English educational broadcast archives, with original publisher MP3 bytes, derived 16 kHz mono PCM WAV clips, automatic transcripts, word timestamps and per-episode speaker IDs. Collection was initialized on 2026-10-02; source publication dates are preserved separately. The initial pilot recordings were published in 2025; the historical expansion includes older source publication dates. This is an independent corpus made using the processing approach in mit-ccc/audio-scraper, from the group behind RadioTalk. It does not contain recovered original 2018–2019 RadioTalk recordings, and is not an exact replication of its Kaldi-era system, station sampling or talk-radio…
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication. - ⌨ Keyboard shortcuts for efficiency 1. Create a transcript CSV file with your content: 2. Start the Flask application: 3. Access the interface: 1. Authentication 2. Session Setup - Click "Start Session" 3. Recording - Use on-screen controls or keyboard shortcuts: - R: Start recording / Stop recording - Space: Play recording - Enter: Save…
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication. - ⌨ Keyboard shortcuts for efficiency 1. Create a transcript CSV file with your content: 2. Start the Flask application: 3. Access the interface: 1. Authentication 2. Session Setup - Click "Start Session" 3. Recording - Use on-screen controls or keyboard shortcuts: - R: Start recording / Stop recording - Space: Play recording - Enter: Save…
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication. - ⌨ Keyboard shortcuts for efficiency 1. Create a transcript CSV file with your content: 2. Start the Flask application: 3. Access the interface: 1. Authentication 2. Session Setup - Click "Start Session" 3. Recording - Use on-screen controls or keyboard shortcuts: - R: Start recording / Stop recording - Space: Play recording - Enter: Save…
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication. - ⌨ Keyboard shortcuts for efficiency 1. Create a transcript CSV file with your content: 2. Start the Flask application: 3. Access the interface: 1. Authentication 2. Session Setup - Click "Start Session" 3. Recording - Use on-screen controls or keyboard shortcuts: - R: Start recording / Stop recording - Space: Play recording - Enter: Save…
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication. - ⌨ Keyboard shortcuts for efficiency 1. Create a transcript CSV file with your content: 2. Start the Flask application: 3. Access the interface: 1. Authentication 2. Session Setup - Click "Start Session" 3. Recording - Use on-screen controls or keyboard shortcuts: - R: Start recording / Stop recording - Space: Play recording - Enter: Save…
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference A web-based interface for preparing audio datasets to fine-tune OpenAI's Whisper model. This tool helps in recording, managing, and organizing voice recordings with their corresponding transcriptions, with support for cloud storage and authentication. - ⌨ Keyboard shortcuts for efficiency 1. Create a transcript CSV file with your content: 2. Start the Flask application: 3. Access the interface: 1. Authentication 2. Session Setup - Click "Start Session" 3. Recording - Use on-screen controls or keyboard shortcuts: - R: Start recording / Stop recording - Space: Play recording - Enter: Save…
The Burmese Daily Dialogue Corpus ( BDDC ) is a synthetic speech dataset based on everyday conversational speech in Myanmar (Burmese). All 13 speakers in the dataset are synthetic Burmese speakers, and no human individual's voice recordings are used. Each speaker was independently designed to have different Speaking Styles and Voice Characteristics. The two test speakers are unseen speakers who are not included in the training or validation sets. To support speaker-independent evaluation, two speakers who are not included in the training or validation sets are reserved exclusively for the test set. This setup allows models to be evaluated on unseen-speaker generalization. In the…
Burmese Daily Dialogue Corpus (BDDC) သည် မြန်မာဘာသာဖြင့် နေ့စဉ်စကားပြောပုံစံများကို အခြေခံထားသော Synthetic Speech Dataset ဖြစ်ပါသည်။ ဤ dataset တွင် ပါဝင်သော Speaker ၁၃ ဦးသည် Synthetic Burmese Speakers များဖြစ်ပြီး မည်သည့် လူသားတစ်ဦးတစ်ယောက်၏ အသံကိုမျှ အသုံးပြုထားခြင်း မရှိပါ။ Speaker တစ်ဦးချင်းစီတွင် စကားပြောပုံ (Speaking Style) နှင့် အသံအရည်အသွေး (Voice Characteristics) မတူညီစေရန် ဖန်တီးထားပါသည်။ လူသားအသံ recording များကို voice cloning သို့မဟုတ် voice conversion ပြုလုပ်ရန် အသုံးပြုခြင်းမရှိသကဲ့သို့ အခြား dataset များမှ အသံများကိုလည်း ထုတ်ယူအသုံးပြုထားခြင်း မရှိပါ။ Test speakers ၂ ဦးသည် Training နှင့် Validation တွင် မပါဝင်သော unseen speakers များဖြစ်ပါသည်။ Speaker-independent evaluation…
IELTS Speaking test recordings from Amanafif554/IELTS-WAV-FILE-P3 (revision 8c68bdf8), each with an automatic Examiner/Candidate transcription and the band score stated in the video title. data/.wav + data/metadata.jsonl – the dataset itself transcriptions/.txt – the transcription of each WAV transcriptions/.json – per-file details: duration, sample rate, channels, size, sha256, band score and how it was found, title, YouTube URL, word counts, candidate word share, speaker check, ASR settings and time-stamped segments with speaker labels WhisperX with whisper-large-v3 (English, float16, a prompt containing fillers so that "um"/"uh" are kept), wav2vec2 word alignment and pyannote speaker…
Built from thelarryproject/scotus-unaligned. U.S. Supreme Court oral arguments, October Terms 2010–2025, cut into short single-speaker clips with matching transcript text. Each argument's transcript was aligned to its audio, then split at speaker changes, long pauses, and sentence ends into clips of at most 25 seconds. Contains 395,983 clips (~1,191 hours) from all 1,050 argument sessions, un-pruned. - docket, term, casename, dateargued, pdfurl, pageurl: copied from the argument the clip came from - audio: Arrow audio typed column of (bytes, path); 16 kHz mono FLAC, path is empty - textoriginal: the transcript text for the clip, as written by the court reporter - textnormalized: the same…
Audio-only repackaging of Meta's dataset (improvised and naturalistic subsets), CC-BY-NC 4.0. resampled from 48 kHz float to 24 kHz 16-bit PCM, stored as FLAC. All label fields are kept as published: - vad, transcript: the per-file metadata/.jsonl records. - annotation1PIS, annotation1PR, annotation3PIS, annotation3PR, annotation3PV: the raw annotation JSON (null when not annotated). - batchidx, archiveidx, has flags: from the original filelist.
The data/ar/ partition of espnet/yodas3, repackaged as parquet with the audio inline so it loads without a script: Every row is one long-form recording (median ~5 minutes) with its full transcript. Audio container only. Upstream ships Opus in a WebM/Matroska container, which libsndfile cannot open. Here it is Opus in Ogg, remuxed with ffmpeg -c:a copy — the Opus packets are copied untouched, so the audio is bit-identical to upstream. 48 kHz. Metadata joined. Upstream keeps audio in data/ar/audio/.tar and the 12 metadata columns in data/ar/metadata/.parquet; this version carries both in one row. Dialect columns added (see below). Nothing was filtered, resampled, re-encoded, segmented, or…
Uzbek audiobook recordings with machine-generated, structurally filtered transcripts. - audio/ — the raw full-length recordings, unmodified. - manifest/gold.jsonl — segments that pass every structural check. Training set. - manifest/review.jsonl — segments held back with the reason recorded. Kept, not discarded. - manifest/reject.jsonl — segments failing a hard check. - manifest/SUMMARY.json — counts, hours, and the gate thresholds actually used. audiohf is a path inside this repo, so a segment can be cut directly from audio/. text is the transcript as decoded; textnorm is case-folded and punctuation-stripped with the okina (U+02BB) preserved as a letter and apostrophe variants folded onto…