Describe SQA+SED provenance, duration and grouped metric families

README.md

@@ -1,29 +1,37 @@
----
+<!-- ---
 language: [de, en, fr, ja, ko, zh]
 license: cc-by-nc-4.0
 pretty_name: Emilia ERT
 size_categories:
 - 1M<n<10M
 tags: [audio, speech, speech-quality, speech-metrics, emilia]
----
+--- -->
 
 # Emilia ERT
 
 ## Dataset summary
 
-Emilia ERT contains **9,310,186 speech segments** in six languages, distributed as **666 audio TAR shards** and a single **34-column metadata table**, `metrics.scp`. It packages the entire locally retained `Emilia_complete` collection from a filtering workflow, with existing transcripts and automatic audio metrics.
+Emilia ERT contains **9,310,186 speech segments totaling 26,133.85 hours** in six languages, distributed as **666 audio TAR shards** and a single **34-column metadata table**, `metrics.scp`. The data originate from `emilia_complete` and were selected using the **SQA+SED filtering strategy described in the accompanying paper**, combining speech quality assessment (SQA) and sound event detection (SED). Existing transcripts and automatic audio metrics are included.
 
 `Emilia_complete` identifies the local source collection; it does **not** mean this repository contains the complete upstream release. No additional English/Chinese-only selection was applied during release preparation. This is an independently prepared collection in the AudioCC-Lab namespace, not an official Amphion release.
 
-| Language | Code | Segments | TAR shards |
-| --- | --- | ---: | ---: |
-| German | de | 166,600 | 10 |
-| English | en | 4,619,380 | 333 |
-| French | fr | 78,751 | 12 |
-| Japanese | ja | 203,264 | 27 |
-| Korean | ko | 11,640 | 1 |
-| Chinese | zh | 4,230,551 | 283 |
-| **Total** | | **9,310,186** | **666** |
+| Language | Code | Segments | Duration (hours) | TAR shards |
+| --- | --- | ---: | ---: | ---: |
+| German | de | 166,600 | 433.62 | 10 |
+| English | en | 4,619,380 | 13,370.38 | 333 |
+| French | fr | 78,751 | 191.89 | 12 |
+| Japanese | ja | 203,264 | 391.90 | 27 |
+| Korean | ko | 11,640 | 30.20 | 1 |
+| Chinese | zh | 4,230,551 | 11,715.86 | 283 |
+| **Total** | | **9,310,186** | **26,133.85** | **666** |
+
+Durations are sums of the `length(s)` values in the released `metrics.scp`, converted to hours and rounded to two decimal places. All 9,310,186 records have numeric duration values.
+
+## Data source and filtering
+
+The source collection is `emilia_complete`, derived from the original [Emilia dataset](https://huggingface.co/datasets/amphion/Emilia-Dataset). This release contains the data retained after applying the paper's **SQA+SED** filtering strategy: SQA provides speech-quality measurements, while SED provides sound-event information used in the selection. The source is recorded as `Emilia_complete` in the `datasets` column. Emilia-YODAS is not included.
+
+Release preparation preserved the already selected audio and its metadata; it did not rerun the filtering strategy. The duration statistics above describe this filtered release.
 
 Audio archives occupy 665,879,367,680 bytes; `metrics.scp` occupies 6,134,693,898 bytes. No held-out evaluation split is provided. Define appropriate speaker/source-disjoint splits for experiments.
 
@@ -51,7 +59,7 @@ The components are **TAR path : byte offset : byte length**. The offset points t
 
 ## Metadata fields
 
-Types describe valid values; all cells are stored as text. Handle missing or nonnumeric markers explicitly rather than converting them to zero.
+Types describe valid values; all cells are stored as text. Handle missing or nonnumeric markers explicitly rather than converting them to zero. Related outputs are grouped into a single metric entry below; column numbers refer to the unchanged 34-column SCP schema.
 
 | # | Field | Type / unit | Meaning |
 | ---: | --- | --- | --- |
@@ -66,9 +74,7 @@ Types describe valid values; all cells are stored as text. Handle missing or non
 | 9 | `confidence` | float, score | Mean of available WhisperX aligned-word scores in the inspected implementation; not a calibrated probability of transcript correctness. |
 | 10 | `dnsmos` | float, predicted MOS | DNSMOS overall quality output (`DNSMOS_OVRL`). |
 | 11 | `scoreq` | float, model score | SCOREQ speech-quality estimate on the model's native scale. |
-| 12 | `squim_pesq` | float, predicted score | SQUIM estimate of PESQ without a clean reference. |
-| 13 | `squim_sisdr` | float, predicted dB | SQUIM estimate of scale-invariant signal-to-distortion ratio (SI-SDR). |
-| 14 | `squim_stoi` | float, predicted score | SQUIM estimate of short-time objective intelligibility (STOI). |
+| 12–14 | **SQUIM:** `squim_pesq`, `squim_sisdr`, `squim_stoi` | float; PESQ score, SI-SDR in dB, STOI score | Non-intrusive estimates of perceptual speech quality (PESQ), scale-invariant signal-to-distortion ratio (SI-SDR), and short-time objective intelligibility (STOI), respectively. These are model predictions without a clean reference. |
 | 15 | `srmr` | float, ratio | Speech-to-reverberation modulation energy ratio, a non-intrusive reverberation-related measure. |
 | 16 | `vqscore` | float, model score | VQScore non-intrusive speech-quality estimate. |
 | 17 | `wadasnr` | float, dB | Signal-to-noise ratio estimated using waveform amplitude distribution analysis (WADA-SNR). |
@@ -76,23 +82,14 @@ Types describe valid values; all cells are stored as text. Handle missing or non
 | 19 | `utmos` | float, predicted MOS | UTMOS speech-quality estimate. |
 | 20 | `dnsmos_pro` | float, model score | DNSMOS Pro quality output retained from the local scoring workflow. |
 | 21 | `nisqa` | float, predicted MOS | NISQA overall quality estimate (`NISQA_MOS`). |
-| 22 | `sigmos_ovrl` | float, predicted rating | SIGMOS overall quality rating. |
-| 23 | `sigmos_col` | float, predicted rating | SIGMOS coloration-related quality rating. |
-| 24 | `sigmos_disc` | float, predicted rating | SIGMOS discontinuity-related quality rating. |
-| 25 | `sigmos_loud` | float, predicted rating | SIGMOS loudness-related quality rating. |
-| 26 | `sigmos_noise` | float, predicted rating | SIGMOS background-noise-related quality rating. |
-| 27 | `sigmos_reverb` | float, predicted rating | SIGMOS reverberation-related quality rating. |
-| 28 | `sigmos_sig` | float, predicted rating | SIGMOS speech-signal quality rating. |
+| 22–28 | **SIGMOS:** `sigmos_ovrl`, `sigmos_col`, `sigmos_disc`, `sigmos_loud`, `sigmos_noise`, `sigmos_reverb`, `sigmos_sig` | float, predicted quality ratings | Ratings for overall quality, coloration, discontinuity, loudness, background noise, reverberation, and speech-signal quality, respectively. These are predicted quality dimensions, not physical impairment measurements. |
 | 29 | `ced` | float, probability score | Maximum CED probability over the workflow's selected noise/event classes (AudioSet indices 72 onward). |
 | 30 | `flexsed` | float, probability score | Maximum FlexSED probability over selected noise/event prompts and time frames in the inspected workflow. |
 | 31 | `panns` | float, probability score | Maximum PANNs clip-level probability over selected classes at AudioSet indices 72 onward. |
 | 32 | `beats` | string | BEATs highest-scoring selected non-speech event label followed by its score, e.g. `Inside, small room 0.0067`. Split from the right to extract the final numeric score. |
-| 33 | `speaker_count` | integer, count | Number of Sortformer speaker streams meeting minimum active-duration criteria; not a verified count of distinct people. |
-| 34 | `overlap_ratio` | float, fraction | Fraction of analyzed frames with at least two active speaker streams after thresholding and short-run removal. The denominator includes all analyzed frames, not only speech. |
+| 33–34 | **Sortformer:** `speaker_count`, `overlap_ratio` | integer count; float fraction | Both are derived from Sortformer diarization outputs. `speaker_count` counts speaker streams meeting minimum active-duration criteria. `overlap_ratio` is the fraction of all analyzed frames with at least two active streams after thresholding and short-run removal; its denominator is not restricted to speech frames. These are automatic estimates. |
 
-Meanings were checked against available local assembly and scoring scripts. Exact historical checkpoints, command-line settings, and per-row generation versions are not fully recorded. Current script defaults are not guaranteed historical settings.
 
-MOS fields are model predictions, not listening-test ratings. SQUIM outputs are predictions rather than reference-based measurements. SIGMOS dimensions describe quality, not physical noise power or reverberation time. The SED class selection is a workflow convention; event scores are not human-confirmed noise labels. Different models' scores are not interchangeable, and this release defines no universal quality threshold.
 
 ## Usage
 
@@ -119,32 +116,9 @@ print(row["uid"], row["transcript"])
 # waveform, sample_rate = sf.read(io.BytesIO(audio_bytes))
 ```
 
-This custom TAR/SCP layout requires a table-driven reader. An automatic WebDataset loader does not join SCP annotations to archive members. For efficient training, group reads by shard and reuse file handles. `SHA256SUMS` records the SHA-256 of `metrics.scp`.
-
-## Provenance and preparation
-
-The upstream dataset is [Emilia, by Haorui He and collaborators / Amphion](https://huggingface.co/datasets/amphion/Emilia-Dataset). Local records identify the original Emilia branch; no Emilia-YODAS collection was included.
-
-Immediate inputs were the six-language `30wh_filter/data/Emilia_complete` audio collection, its language manifests, and corresponding rows in six `metrics_dataset.*.scp` tables. Preparation consolidated those rows, preserved audio bytes and metric values, and replaced local addresses with relative TAR locators. It did not rerun ASR, scoring, diarization, or filtering. Record IDs, source locators, and all non-path fields were checked during preparation.
 
-The exact upstream download revision, complete earlier filtering history, original media URLs, and per-record rights records are not supplied. Local consistency checks do not establish a complete upstream provenance or permission chain.
 
 ## License and access review
 
-Reviewed on **2026-09-18**. The [upstream card and access terms](https://huggingface.co/datasets/amphion/Emilia-Dataset) distinguish original Emilia (**CC BY-NC 4.0**) from Emilia-YODAS (**CC BY 4.0**). This repository retains `cc-by-nc-4.0`; the combined upstream repository's license badge should not be used to relabel original Emilia. Upstream states that recording copyright remains with the original owners.
-
-[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/legalcode.en) permits noncommercial sharing and adaptation within the licensor's rights, subject to attribution, a license reference, retention of supplied notices, and indication of changes. It does not independently clear third-party rights. See [LICENSE](LICENSE) for the attribution notice. This repackaging claims no upstream endorsement or new ownership of recordings.
-
-**Unresolved release condition:** upstream access terms require associates or colleagues receiving access to agree to those terms first. Available records do not establish how unrestricted public mirroring satisfies this condition or document permission for that distribution arrangement. README text alone does not enforce acceptance. Compliance of unrestricted redistribution is therefore **not confirmed**. The maintainer should obtain upstream clarification/authorization and establish the appropriate access arrangement before treating this release as cleared for unrestricted distribution. Gating alone would not establish recording-level rights.
-
-The license label records the upstream designation, not certification that every recording or annotation has been cleared. Provenance and licensing of annotation-generating models/checkpoints are not fully documented here either.
-
-## Limitations and responsible use
-
-Languages are strongly imbalanced. Automatic transcripts, speaker estimates, and quality/event scores can be wrong and may behave differently across languages and recording conditions. No guarantee of absence of personal information or sensitive content is made. Evaluate suitability and rights for the intended use; metadata scores are not human approval or authorization to impersonate speakers.
-
-For dataset questions or rights concerns, contact maintainers through this repository's Community page. Avoid posting personal information publicly.
-
-## Citation
 
-Credit the original Emilia authors and cite their [2024 paper](https://arxiv.org/abs/2407.05361) and, where relevant, the [2025 extended paper](https://arxiv.org/abs/2501.15907). Also identify this repository and the commit used to distinguish its local selection and added metadata from upstream Emilia.
+[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/legalcode.en) permits noncommercial sharing and adaptation within the licensor's rights, subject to attribution, a license reference, retention of supplied notices, and indication of changes. It does not independently clear third-party rights. See [LICENSE](LICENSE) for the attribution notice.