--- dataset_info: - config_name: dutch features: - name: audio dtype: audio: sampling_rate: 48000 - name: key dtype: string - name: language dtype: string - name: source_split dtype: string - name: source_shard dtype: string - name: transcript dtype: string - name: speaker dtype: string - name: meta_json dtype: string - name: duration dtype: float32 - name: sampling_rate dtype: int32 - name: bak dtype: float32 - name: sig dtype: float32 - name: ovrl dtype: float32 - name: p808 dtype: float32 splits: - name: train num_bytes: 7365724100 num_examples: 21018 download_size: 7357400224 dataset_size: 7365724100 - config_name: french features: - name: audio dtype: audio: sampling_rate: 48000 - name: key dtype: string - name: language dtype: string - name: source_split dtype: string - name: source_shard dtype: string - name: transcript dtype: string - name: speaker dtype: string - name: meta_json dtype: string - name: duration dtype: float32 - name: sampling_rate dtype: int32 - name: bak dtype: float32 - name: sig dtype: float32 - name: ovrl dtype: float32 - name: p808 dtype: float32 splits: - name: train num_bytes: 13164833919 num_examples: 36942 download_size: 13149069012 dataset_size: 13164833919 - config_name: german features: - name: audio dtype: audio: sampling_rate: 48000 - name: key dtype: string - name: language dtype: string - name: source_split dtype: string - name: source_shard dtype: string - name: transcript dtype: string - name: speaker dtype: string - name: meta_json dtype: string - name: duration dtype: float32 - name: sampling_rate dtype: int32 - name: bak dtype: float32 - name: sig dtype: float32 - name: ovrl dtype: float32 - name: p808 dtype: float32 splits: - name: train num_bytes: 22184970687 num_examples: 61478 download_size: 22160047323 dataset_size: 22184970687 - config_name: italian features: - name: audio dtype: audio: sampling_rate: 48000 - name: key dtype: string - name: language dtype: string - name: source_split dtype: string - name: source_shard dtype: string - name: transcript dtype: string - name: speaker dtype: string - name: meta_json dtype: string - name: duration dtype: float32 - name: sampling_rate dtype: int32 - name: bak dtype: float32 - name: sig dtype: float32 - name: ovrl dtype: float32 - name: p808 dtype: float32 splits: - name: train num_bytes: 3797822305 num_examples: 10570 download_size: 3793528383 dataset_size: 3797822305 - config_name: polish features: - name: audio dtype: audio: sampling_rate: 48000 - name: key dtype: string - name: language dtype: string - name: source_split dtype: string - name: source_shard dtype: string - name: transcript dtype: string - name: speaker dtype: string - name: meta_json dtype: string - name: duration dtype: float32 - name: sampling_rate dtype: int32 - name: bak dtype: float32 - name: sig dtype: float32 - name: ovrl dtype: float32 - name: p808 dtype: float32 splits: - name: train num_bytes: 2074543715 num_examples: 5864 download_size: 2072195570 dataset_size: 2074543715 - config_name: portuguese features: - name: audio dtype: audio: sampling_rate: 48000 - name: key dtype: string - name: language dtype: string - name: source_split dtype: string - name: source_shard dtype: string - name: transcript dtype: string - name: speaker dtype: string - name: meta_json dtype: string - name: duration dtype: float32 - name: sampling_rate dtype: int32 - name: bak dtype: float32 - name: sig dtype: float32 - name: ovrl dtype: float32 - name: p808 dtype: float32 splits: - name: train num_bytes: 3076946267 num_examples: 8389 download_size: 3073657132 dataset_size: 3076946267 - config_name: spanish features: - name: audio dtype: audio: sampling_rate: 48000 - name: key dtype: string - name: language dtype: string - name: source_split dtype: string - name: source_shard dtype: string - name: transcript dtype: string - name: speaker dtype: string - name: meta_json dtype: string - name: duration dtype: float32 - name: sampling_rate dtype: int32 - name: bak dtype: float32 - name: sig dtype: float32 - name: ovrl dtype: float32 - name: p808 dtype: float32 splits: - name: train num_bytes: 9310737326 num_examples: 25965 download_size: 9300259828 dataset_size: 9310737326 configs: - config_name: dutch data_files: - split: train path: dutch/train-* - config_name: french data_files: - split: train path: french/train-* - config_name: german data_files: - split: train path: german/train-* - config_name: italian data_files: - split: train path: italian/train-* - config_name: polish data_files: - split: train path: polish/train-* - config_name: portuguese data_files: - split: train path: portuguese/train-* - config_name: spanish data_files: - split: train path: spanish/train-* --- # MLS-Sidon-HQ The **top-DNSMOS slice** of [`sarulab-speech/mls_sidon`](https://huggingface.co/datasets/sarulab-speech/mls_sidon) — Multilingual LibriSpeech (LibriVox) restored to 48 kHz by [`sarulab-speech/sidon-v0.1`](https://huggingface.co/sarulab-speech/sidon-v0.1), then scored clip-by-clip with **DNSMOS P.835** and cut down to only the cleanest utterances. All **1,473,249** source clips (6,143 h, 0.91 TB of FLAC) were scored; **170,226** (11.6%, **702 h**) passed and are published here. ## Filter DNSMOS P.835 (`speechmos`, ONNX) on a single centred 10 s window at 16 kHz, peak-normalised. A clip is kept only when **every** gate passes: | metric | threshold | |---|---| | `ovrl` — overall MOS | ≥ 3.5 | | `sig` — speech-signal quality | ≥ 3.7 | | `bak` — background-noise quality | ≥ 4.15 | | duration | 1–40 s | These are strict on purpose. The source is already restored, so `bak` sits near the DNSMOS ceiling (~4.2) across the board and the discriminating axes are `sig` and `ovrl`; the bar above sits roughly at the source's 96th–99th percentile. ## Layout One **parquet config per language**, same shape as [`Scicom-intl/TTS-Clean44k`](https://huggingface.co/datasets/Scicom-intl/TTS-Clean44k). Audio is **mp3, 192 kbps mono, 48 kHz**. | column | type | | |---|---|---| | `audio` | `Audio(48000)` | mp3 bytes | | `key` | string | WebDataset key in the source shard | | `language` | string | config name | | `source_split` | string | `train` / `validation` / `test` / … in the source | | `source_shard` | string | exact source shard this came from | | `transcript` | string | CV22 `sentence` / MLS `transcript` | | `speaker` | string | CV22 `client_id` / MLS `speaker_id` | | `meta_json` | string | the source metadata verbatim (CV22: age, gender, accent, locale, votes; MLS: chapter_id, original_path, begin/end time) | | `duration` | float32 | seconds | | `bak` `sig` `ovrl` `p808` | float32 | this clip's DNSMOS P.835 scores | ```python from datasets import load_dataset ds = load_dataset("Scicom-intl/MLS-Sidon-HQ", "dutch", split="train") ``` | config | clips | hours | keep rate | mean ovrl-gated source | |---|---|---|---|---| | `german` | 61,478 | 255.6 | 12.9% | 1995 h scored | | `french` | 36,942 | 151.6 | 14.0% | 1097 h scored | | `spanish` | 25,965 | 107.3 | 11.5% | 938 h scored | | `dutch` | 21,018 | 84.9 | 5.5% | 1580 h scored | | `italian` | 10,570 | 43.8 | 17.0% | 258 h scored | | `portuguese` | 8,389 | 35.5 | 21.4% | 168 h scored | | `polish` | 5,864 | 23.9 | 22.5% | 108 h scored | | **total (7 configs)** | **170,226** | **702.4** | **11.6%** | **6143 h scored** | ## Caveats - **Lossy.** Source audio is 48 kHz FLAC; this republishes it as mp3. Use the source repo if you need bit-exact audio. - **Restored, not native.** Every clip has been through sidon-v0.1, so it carries that model's characteristics. DNSMOS rates restored audio generously — a high `ovrl` here means "clean and artefact-free *for a restored clip*", which is not the same as native studio capture. - **DNSMOS is non-intrusive and imperfect.** It scores perceptual cleanliness, not transcript accuracy or speaker consistency. ## Provenance & licence `cc-by-4.0`, inherited from Multilingual LibriSpeech (LibriVox) via `sarulab-speech/mls_sidon`. This repo only *selects* and re-encodes; no audio was otherwise altered. Please cite MLS and Sidon.