ftshijt commited on
Commit
0055db1
·
verified ·
1 Parent(s): d189d6b

Publish validated Speech-DRAME paper metadata and documentation

Browse files
.gitattributes CHANGED
@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ metadata/realism/train.jsonl filter=lfs diff=lfs merge=lfs -text
LICENSE.md ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Dataset license
2
+
3
+ Author-controlled annotations, generated context, captions, splits, statistics, and
4
+ derived assets authorized by the authors for this release are provided under
5
+ Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).
6
+
7
+ https://creativecommons.org/licenses/by-nc/4.0/
8
+ https://creativecommons.org/licenses/by-nc/4.0/legalcode.en
9
+
10
+ Underlying third-party works, model weights, software, and service outputs retain
11
+ their applicable terms and attribution. This grant does not relicense rights the
12
+ authors do not hold. No original source waveforms or vocoder weights are included.
13
+ The BigVGAN-v2 vocoder is separately MIT-licensed. Feature conversion alone does
14
+ not provide a legal basis for redistribution. See SOURCE_NOTES.md.
15
+
16
+ No additional research-only restriction is added to this CC BY-NC license.
README.md CHANGED
@@ -1,15 +1,157 @@
1
  ---
2
  license: cc-by-nc-4.0
3
- pretty_name: "DRAME-EvalBench — Metadata and Mel Features"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  ---
5
 
6
- # DRAME-EvalBench — Metadata and Mel Features
7
 
8
- **Dataset release in preparation. No data files are available yet.**
 
 
 
 
9
 
10
- The planned release contains authorized benchmark metadata and mel features. Pre-rendered vocoder reconstructions will be hosted in the companion [ftshijt dataset](https://huggingface.co/datasets/ftshijt/speech-drame-evalbench).
11
 
12
- Only the subset covered by the source-specific release grants will be included. Additional internal development data and original source waveforms will not be distributed. Approved contributions will use [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), with applicable source notices retained. Final counts and documentation will be added with the data release.
 
13
 
14
- Evaluation code: [Anuttacon/speech_drame](https://github.com/Anuttacon/speech_drame).
15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: cc-by-nc-4.0
3
+ language:
4
+ - en
5
+ - zh
6
+ pretty_name: "Speech-DRAME-EvalBench — Metadata and Mel Features"
7
+ tags:
8
+ - speech
9
+ - role-play
10
+ - human-evaluation
11
+ configs:
12
+ - config_name: archetype
13
+ data_files:
14
+ - split: test
15
+ path: metadata/archetype/test.jsonl
16
+ - split: train
17
+ path: metadata/archetype/train.jsonl
18
+ - config_name: realism
19
+ data_files:
20
+ - split: benchmark
21
+ path: metadata/realism/benchmark.jsonl
22
+ - split: realrec_adaptation
23
+ path: metadata/realism/realrec_adaptation.jsonl
24
+ - split: realrec_clean_test
25
+ path: metadata/realism/realrec_clean_test.jsonl
26
+ - split: test_base
27
+ path: metadata/realism/test_base.jsonl
28
+ - split: train
29
+ path: metadata/realism/train.jsonl
30
  ---
31
 
32
+ # Speech-DRAME-EvalBench — Metadata and Mel Features
33
 
34
+ This release contains the **23,686 annotated experimental records used in the paper**
35
+ [Speech-DRAME: Human-Aligned Evaluation for Speech Role-Play](https://openreview.net/forum?id=m9IP9ECrk2).
36
+ It includes **8,272 archetype** and **15,414 realism** records. Additional internal
37
+ development data, the 1,250 unannotated future generation prompts, model checkpoints,
38
+ and original source waveforms are excluded.
39
 
40
+ This variant includes **metadata and compressed mel features only**. It contains no FLAC/WAV files or reconstructed-audio payloads.
41
 
42
+ The [companion ftshijt dataset](https://huggingface.co/datasets/ftshijt/speech-drame-evalbench)
43
+ contains the same record IDs, split assignments, annotations, and mel features, with pre-rendered reconstructed audio.
44
 
45
+ ## Splits
46
 
47
+ | Perspective | Split | Records | Meaning |
48
+ |---|---|---:|---|
49
+ | Archetype | train | 6,773 | Evaluator training |
50
+ | Archetype | test | 1,499 | Held-out role × scene combinations |
51
+ | Realism | train | 10,990 | Base evaluator training |
52
+ | Realism | test_base | 2,000 | Benchmark-distribution test; 210 positives, 832 role and 958 context negatives |
53
+ | Realism | realrec_adaptation | 1,080 | Additional real-recording training data |
54
+ | Realism | realrec_clean_test | 1,084 | Clean held-out human acting; speakers disjoint from adaptation |
55
+ | Realism | benchmark | 260 | Separate delivery-mismatch benchmark; 13 TTS systems × 20 Mandarin scenes |
56
+
57
+ The real-recording pool is partitioned, not duplicated: 1,080 + 1,084 = 2,164.
58
+ The final adapted realism evaluator uses 10,990 + 1,080 = 12,070 training records.
59
+ The real-recording and delivery-benchmark splits are Mandarin-only. Archetype and
60
+ the base realism splits contain English and Mandarin. Base realism is mixed
61
+ human/TTS audio. Clip disjointness does not imply character disjointness.
62
+
63
+ ## Download and load
64
+
65
+ Metadata are JSON Lines. Asset files are bundled in uncompressed tar shards of
66
+ approximately 256 MiB; each asset entry names both its archive and extraction path.
67
+ Download the repository and extract only the desired assets:
68
+
69
+ ```python
70
+ from huggingface_hub import snapshot_download
71
+ from pathlib import Path
72
+ import json
73
+ import tarfile
74
+ import numpy as np
75
+
76
+ root = Path(snapshot_download("Anuttacon/speech-drame-evalbench", repo_type="dataset"))
77
+ # Use Python 3.12+ for the safe tar extraction filter.
78
+ extracted = Path("speech-drame-extracted")
79
+ extracted.mkdir(exist_ok=True)
80
+ for shard in sorted((root / "assets").rglob("*.tar")):
81
+ with tarfile.open(shard) as archive:
82
+ archive.extractall(extracted, filter="data")
83
+
84
+ records = [json.loads(line) for line in
85
+ (root / "metadata/realism/realrec_clean_test.jsonl").read_text().splitlines()]
86
+ mel = np.load(extracted / records[0]["audio"]["spectrogram"]["path"],
87
+ allow_pickle=False)["mel"]
88
+ ```
89
+
90
+ Each tar contains regular files only. `manifests/assets.jsonl` records the
91
+ SHA-256 and archive membership of every asset; `manifests/files.sha256` covers
92
+ all release files except itself. Metadata can also be loaded using the `archetype`
93
+ or `realism` configuration with the Hugging Face `datasets` library.
94
+
95
+ ## Annotation and provenance notes
96
+
97
+ Archetype annotations include Audio Quality, Human-likeness, Appropriateness, and
98
+ Content Pass. Realism retains the eleven annotation sub-dimensions; the paper's
99
+ evaluator predicts ten, excluding Semantic Match as a separate output. Per-rater
100
+ arrays and missing/gated values are preserved; missing ratings remain JSON null.
101
+ Use masking for the paper's primary correlations. The base realism training set
102
+ contains 31 two-rater records; evaluation records have at least three raters.
103
+
104
+ Positive/negative construction is separate from audio provenance. Role/context
105
+ negatives change text and retain their anchor clip's audio, which may be human
106
+ or TTS; delivery negatives resynthesize audio. In `test_base`, each served record
107
+ has a distinct source clip, with no second matching positive record in that set.
108
+
109
+ The old `is_real_audio` field was unreliable and is preserved only as
110
+ `legacy_is_real_audio`, **not as ground-truth provenance**. `source_type` is `human`
111
+ for the real-recording pool and `tts` for archetype/delivery-benchmark records;
112
+ it is `unverified` for the mixed base realism records pending a corrected
113
+ clip-level provenance mapping. Existing nonempty source attributes are retained;
114
+ missing source information is not fabricated. Every distributed FLAC, where
115
+ present, is a vocoder reconstruction regardless of source type.
116
+
117
+ ## Reproduction and limitations
118
+
119
+ The exact six canonical experimental metadata inputs were checked against the
120
+ paper inventory. Record IDs, annotations, and scientific context are retained;
121
+ split names, asset references, and provenance warnings have been normalized for
122
+ public use. Public metadata input hashes and counts are in `release.json`.
123
+ Mel values, dimensions, FLAC headers, and all derived asset checksums were checked
124
+ before publication. These files reproduce release-native masked evaluation;
125
+ legacy analyses that also floor model predictions require additional saved
126
+ prediction artifacts. Code and model weights are separate releases.
127
+
128
+ `vocoder/` contains the original mel frontend settings (100 bands, 24 kHz,
129
+ 1024-point FFT, 256-sample hop) and the BigVGAN-v2 model identifier. Mel features
130
+ and reconstructions retain voice-identity information and do not eliminate
131
+ re-identification or voice-cloning risk. Scores reflect the bilingual annotator
132
+ populations and can encode stereotypes. The paper reports strong benchmark
133
+ alignment but poor synthetic-to-authentic transfer; this evaluator is not a
134
+ validated general-purpose judge of human acting.
135
+
136
+ ## License and citation
137
+
138
+ Author-controlled contributions are released under **CC BY-NC 4.0**. Underlying
139
+ source rights, attribution, and third-party model/service terms remain with their
140
+ owners; this release does not relicense third-party works. Feature conversion
141
+ does not itself establish redistribution rights. See `LICENSE.md` and
142
+ `SOURCE_NOTES.md`. No additional research-only restriction is imposed.
143
+
144
+ - [Paper](https://openreview.net/forum?id=m9IP9ECrk2)
145
+ - [Project](https://ftshijt.github.io/speech_drame/)
146
+ - [Evaluation code](https://github.com/Anuttacon/speech_drame)
147
+
148
+ ```bibtex
149
+ @article{shi2026speechdrame,
150
+ title={Speech-DRAME: Human-Aligned Evaluation for Speech Role-Play},
151
+ author={Shi, Jiatong and Han, Jionghao and Lu, Yichen and Pascual, Santiago and
152
+ Wu, Pengfei and Cui, Chenye and Watanabe, Shinji and Weng, Chao and Zhou, Cong},
153
+ journal={Transactions on Machine Learning Research},
154
+ year={2026},
155
+ url={https://openreview.net/forum?id=m9IP9ECrk2}
156
+ }
157
+ ```
SOURCE_NOTES.md ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Source and representation notes
2
+
3
+ This public package is restricted to the six canonical annotated paper datasets
4
+ and their referenced mel features.
5
+ Additional development datasets and original source waveforms are excluded.
6
+
7
+ Realism source material includes NCSSD, media/game-derived speech, and original
8
+ recordings; archetype material comprises generated speech from the systems named
9
+ in `sfnd.system`. The original recording pool contains professional and amateur
10
+ acting. Source credits in `original_id`, profiles, and available metadata are
11
+ retained. Full source and system references are in the paper:
12
+ https://openreview.net/forum?id=m9IP9ECrk2
13
+
14
+ The source metadata in the prepared mixed-realism records is incomplete. Empty
15
+ source objects remain empty; no source license or consent document is fabricated.
16
+ `legacy_is_real_audio` is unreliable and must not be used to label provenance.
17
+ Consult the paper and original corpus/provider terms before reuse. Original
18
+ recording consent and third-party permissions are not transferable permissions
19
+ to clone voices or identify speakers. Mel features and reconstructions remain
20
+ biometric; pseudonymization does not remove that risk.
21
+
22
+ The authors selected the paper datasets for this public release. The CC BY-NC
23
+ grant applies only to contributions they are authorized to license; underlying
24
+ rights and third-party terms remain with the relevant holders. No claim is made
25
+ that conversion to a mel or a vocoder reconstruction removes those rights.
manifests/assets.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
manifests/files.sha256 ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ fae10b8bda57e5fac5b229ce1a2b2ae21aaf78f566572e10c7b858b3dff2d5f2 LICENSE.md
2
+ ae9714f72ee9e333d612b80110332c8e7541a2ad8e773ae217b4ea466c46d93a README.md
3
+ a0b32f5ec8ed762674091e53b725317c6e72679d36332e94e7f0c638a026eca3 SOURCE_NOTES.md
4
+ 0662c82ee12691c71810ada3cae25f9dc007a8705e8be2164528ab3f2b9e99cf assets/archetype/mel/test-000.tar
5
+ 4a4a4c79d2a76477bd8c4ea89db0aff5525ddb434eaa700508b894b70b5ed2c4 assets/archetype/mel/test-001.tar
6
+ d0d376aa1277f441cee9212bce371c4b949cdaa09385924cb40bb16e642cc4aa assets/archetype/mel/train-000.tar
7
+ e207f8fadeb0d270d20b48b18d127b10d86e7f14caba039b95f314a134ce7f38 assets/archetype/mel/train-001.tar
8
+ b4114999af7153e044283139cee771a90d1118bae036b9d6ab1fea9d4dbc13ac assets/archetype/mel/train-002.tar
9
+ c9026a4e9eee7baa74c94f77ae2af4b14b5cf89e9a9e4e638a82c8b025233631 assets/archetype/mel/train-003.tar
10
+ 016dc0adb3ffa9928e69031b7f1c89a84f61a47324cffb215e661ead00927585 assets/archetype/mel/train-004.tar
11
+ 9702a0953169eb2ce7aa682dd942d30916e209414da3911797031c076b6baccf assets/archetype/mel/train-005.tar
12
+ d7644fc94b16f4539c1218f9557f67c8d8c82638b62049faa956c5753e2a4e75 assets/archetype/mel/train-006.tar
13
+ bfef5243fb2560c1b6f3aa6d434b41edcd3c1cba066c6e36c86545515689de15 assets/realism/mel/benchmark-000.tar
14
+ 147f92de7daa3acac8938604adfd783549f109b53fbd128c54a79490e7f6853e assets/realism/mel/realrec_adaptation-000.tar
15
+ 0c463aedc4bc048bd95e43490a45eb9577466d5c5c54027f90bbdcb64b6fb902 assets/realism/mel/realrec_clean_test-000.tar
16
+ 6a0f13fbc89da1b80efebdd8faf399c2d9244fa8f4b561cf9f7479c54eaccf0a assets/realism/mel/test_base-000.tar
17
+ 3bd6252aa66e2f2774d7e01fc2fb1b01bbf1dac21e919098f07bfab791ecaac5 assets/realism/mel/test_base-001.tar
18
+ 55a5bf32dd205de473c0a0ed62ac701945395e0bc9dfa51d9a1a8245d64eadbc assets/realism/mel/train-000.tar
19
+ 0e85bbf6728dad9235be2f21630a0603a843d21cda5ff83739ace36b8e147d65 assets/realism/mel/train-001.tar
20
+ 6ac0cd737a766bb3234fba959a8e8ba3ef8b0b4082f60889e6874c5f215d823f assets/realism/mel/train-002.tar
21
+ 07a75246c05852b588d0cde838048e5504a269ff9852c22759892eda61c556c9 assets/realism/mel/train-003.tar
22
+ fbdb0bbacc8c8991926b9e3d89fad4d64d3a52705d09e5eca6ad7af47010668c assets/realism/mel/train-004.tar
23
+ 9bdac8aefa790cde1ed72596d71fdb55b4b8dd8d9ea698b12d0b2db2782fc714 assets/realism/mel/train-005.tar
24
+ 06b92a69c8b0e911122abf2cba8495045169d5ea75c4058deadabd0d7049d504 assets/realism/mel/train-006.tar
25
+ 525bc31b89134a8e0104ff3f8124e36a5709eb2abf9ca726e16ce54bb4235787 assets/realism/mel/train-007.tar
26
+ a1d3656ef4129c25e71f5e67b5c49e05f0cd9b5bfe4099824e01235047372d50 assets/realism/mel/train-008.tar
27
+ 46158f5d32bef1da6bdddd1c645496ef9e973f34011f9f3d7d62aaafdbfdfced manifests/assets.jsonl
28
+ fcc602aedc249fbb239912ede7843727c04d813277b13e8d8f198acc0756f316 metadata/archetype/test.jsonl
29
+ 012f59ebf6ff19539d54b1b2e653728a26a2786a7b25a7ffe43c80222f153cbe metadata/archetype/train.jsonl
30
+ 54ff1bde9113878ddc4abb886495b949ad31235af333df4e88bb1d309a3c78ab metadata/realism/benchmark.jsonl
31
+ 99ba38638eb1db18b2a3a5652d7071aa449f27125a3b0958f33c2896fad16a29 metadata/realism/realrec_adaptation.jsonl
32
+ e362f6723ceee7703c29a16bc4eab0a093f8e23e473ed9a467fc17c3bd76df52 metadata/realism/realrec_clean_test.jsonl
33
+ ab241f9cabc75acfe5c5a33bb734d104f1511f7c2cbb1514dd464f0cf83e794d metadata/realism/test_base.jsonl
34
+ 992b6385bd817f39965d57b1e2559fe0d749726edfd62d08295e60737b1637b8 metadata/realism/train.jsonl
35
+ ab35330bae5607cd30ca8047b259a61ded4829a158c921d91a70fc8f4d48b1ac release.json
36
+ 083c7033f3b4455e8ff303cd5e6330d096a0f3c38ec975d809b872985ce86444 schema/record.schema.json
37
+ 429a26eead98864925f8dfc071c8b4e447dbac7dc7e95c05dc01d4d0d0a8ca59 vocoder/archetype.json
38
+ 429a26eead98864925f8dfc071c8b4e447dbac7dc7e95c05dc01d4d0d0a8ca59 vocoder/realism.json
metadata/archetype/test.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
metadata/archetype/train.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
metadata/realism/benchmark.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
metadata/realism/realrec_adaptation.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
metadata/realism/realrec_clean_test.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
metadata/realism/test_base.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
metadata/realism/train.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:992b6385bd817f39965d57b1e2559fe0d749726edfd62d08295e60737b1637b8
3
+ size 40874033
release.json ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "2026-10-16",
3
+ "records": 23686,
4
+ "counts": {
5
+ "realism": {
6
+ "train": 10990,
7
+ "test_base": 2000,
8
+ "realrec_adaptation": 1080,
9
+ "realrec_clean_test": 1084,
10
+ "benchmark": 260
11
+ },
12
+ "archetype": {
13
+ "train": 6773,
14
+ "test": 1499
15
+ }
16
+ },
17
+ "repo_id": "Anuttacon/speech-drame-evalbench",
18
+ "includes_reconstructed_audio": false,
19
+ "original_waveforms_included": false,
20
+ "assets": 23686,
21
+ "asset_bytes": 4984392461,
22
+ "canonical_metadata": [
23
+ {
24
+ "perspective": "realism",
25
+ "split": "train",
26
+ "records": 10990,
27
+ "sha256": "dece4e98afe5e9c6254c383c827f494b6d9216f774f7a41c54501dd7c63401bd"
28
+ },
29
+ {
30
+ "perspective": "realism",
31
+ "split": "test_base",
32
+ "records": 2000,
33
+ "sha256": "e87afb90b7c48b7d2e423b48324f13433636ac5b579c60cea2421d0b5e117120"
34
+ },
35
+ {
36
+ "perspective": "realism",
37
+ "split": "test_realrec",
38
+ "records": 2164,
39
+ "sha256": "589cc069ae41061389afad2967fa346802d2f80ccf9ad031a8d9251ebd5df3cd"
40
+ },
41
+ {
42
+ "perspective": "realism",
43
+ "split": "benchmark",
44
+ "records": 260,
45
+ "sha256": "b487364f4a491009e3bdf4788803793749e8d1818712f1d6309878d85803c357"
46
+ },
47
+ {
48
+ "perspective": "archetype",
49
+ "split": "train",
50
+ "records": 6773,
51
+ "sha256": "1f702d02528f6a681ed9a4cb302ff00b53db3e486e94f397e73508bade647e4c"
52
+ },
53
+ {
54
+ "perspective": "archetype",
55
+ "split": "test",
56
+ "records": 1499,
57
+ "sha256": "bc892667334966c19ec78c94c24665e81775d6f1be6c2cba8faf28640ad7f64f"
58
+ }
59
+ ],
60
+ "validation": {
61
+ "source_metadata_hashes": "passed",
62
+ "asset_sha256": "passed",
63
+ "mel_shape_dtype_finite": "passed",
64
+ "flac_headers": "passed",
65
+ "record_schema": "passed",
66
+ "clean_partition": "passed"
67
+ }
68
+ }
schema/record.schema.json ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "type": "object",
3
+ "required": [
4
+ "id",
5
+ "perspective",
6
+ "language",
7
+ "split",
8
+ "annotations",
9
+ "audio",
10
+ "source_type"
11
+ ],
12
+ "properties": {
13
+ "id": {
14
+ "type": "string"
15
+ },
16
+ "perspective": {
17
+ "enum": [
18
+ "realism",
19
+ "archetype"
20
+ ]
21
+ },
22
+ "split": {
23
+ "type": "string"
24
+ },
25
+ "annotations": {
26
+ "type": "object"
27
+ },
28
+ "language": {
29
+ "enum": [
30
+ "en",
31
+ "zh"
32
+ ]
33
+ },
34
+ "source_type": {
35
+ "enum": [
36
+ "human",
37
+ "tts",
38
+ "unverified"
39
+ ]
40
+ },
41
+ "audio": {
42
+ "type": "object",
43
+ "required": [
44
+ "spectrogram"
45
+ ],
46
+ "properties": {
47
+ "spectrogram": {
48
+ "type": "object",
49
+ "required": [
50
+ "path",
51
+ "archive",
52
+ "sha256"
53
+ ]
54
+ }
55
+ }
56
+ }
57
+ }
58
+ }
vocoder/archetype.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "representation": "mel_spectrogram",
3
+ "mel_config": {
4
+ "name": "bigvgan_24k_100band",
5
+ "sampling_rate": 24000,
6
+ "n_fft": 1024,
7
+ "hop_size": 256,
8
+ "win_size": 1024,
9
+ "num_mels": 100,
10
+ "fmin": 0,
11
+ "fmax": 12000
12
+ },
13
+ "dynamic_range_compression": "log(clamp(magnitude, min=1e-5))",
14
+ "stft": {
15
+ "window": "hann",
16
+ "center": false,
17
+ "reflect_pad": 384
18
+ },
19
+ "recommended_vocoder": {
20
+ "primary": "nvidia/bigvgan_v2_24khz_100band_256x",
21
+ "license": "MIT",
22
+ "fast_alternative": "charactr/vocos-mel-24khz (MIT)"
23
+ },
24
+ "note": "Float16 log-mel features. BigVGAN reconstruction is an approximation, not an original waveform."
25
+ }
vocoder/realism.json ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "representation": "mel_spectrogram",
3
+ "mel_config": {
4
+ "name": "bigvgan_24k_100band",
5
+ "sampling_rate": 24000,
6
+ "n_fft": 1024,
7
+ "hop_size": 256,
8
+ "win_size": 1024,
9
+ "num_mels": 100,
10
+ "fmin": 0,
11
+ "fmax": 12000
12
+ },
13
+ "dynamic_range_compression": "log(clamp(magnitude, min=1e-5))",
14
+ "stft": {
15
+ "window": "hann",
16
+ "center": false,
17
+ "reflect_pad": 384
18
+ },
19
+ "recommended_vocoder": {
20
+ "primary": "nvidia/bigvgan_v2_24khz_100band_256x",
21
+ "license": "MIT",
22
+ "fast_alternative": "charactr/vocos-mel-24khz (MIT)"
23
+ },
24
+ "note": "Float16 log-mel features. BigVGAN reconstruction is an approximation, not an original waveform."
25
+ }