owensong commited on
Commit
0c1958a
·
verified ·
1 Parent(s): 9c1e420

Update verified private early-access card

Browse files
README.md CHANGED
@@ -1,22 +1,168 @@
1
  ---
2
- language: en
 
3
  pipeline_tag: text-to-speech
 
4
  license: other
5
  license_name: inflect-preview-evaluation-1.0
6
  license_link: LICENSE.md
7
  inference: false
 
 
 
 
 
 
 
8
  ---
9
- <p align="center"><img src="assets/inflect-mark.svg" width="96" alt="Inflect"></p>
10
 
11
- # Inflect Atom v3: early-access staging draft
12
 
13
- Unpublished private staging; not an announced final release. Six fixed English
14
- voices, no reference-audio adaptation or event tags. The unchanged preview
15
- evaluation license covers only Inflect-specific assets to the extent of the
16
- licensor's rights. It restricts commercial use and hosting without separate
17
- written permission and preserves independent upstream terms and earlier grants.
18
- Read LICENSE.md and THIRD_PARTY_NOTICES.md. No publication or quality approval
19
- is implied by this draft. Final install and model QA are pending.
20
 
21
- Run `python -B synthesize.py --help` for the local CLI and
22
- `python -B synthesize.py --verify` for offline manifest verification.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
  pipeline_tag: text-to-speech
5
+ model_name: Inflect Atom v3
6
  license: other
7
  license_name: inflect-preview-evaluation-1.0
8
  license_link: LICENSE.md
9
  inference: false
10
+ tags:
11
+ - inflect
12
+ - text-to-speech
13
+ - english
14
+ - multi-speaker
15
+ - early-access
16
+ thumbnail: https://huggingface.co/owensong/Inflect-Atom-v3/resolve/main/assets/banner-v3.svg
17
  ---
 
18
 
19
+ <p align="center"><img src="assets/banner-v3.svg" alt="Inflect v3. Mini, Micro and Atom. Local English speech, noncommercial early access." width="100%"></p>
20
 
21
+ <h1 align="center">Inflect Atom v3</h1>
22
+ <p align="center"><strong>A very small synthesis core. A complete speech waveform.</strong><br>Six fixed voices. Local inference. Noncommercial early access.</p>
 
 
 
 
 
23
 
24
+ <p align="center"><a href="#run-locally"><img src="assets/v3-link-run.svg" alt="RUN LOCALLY" height="34"></a>
25
+ <a href="https://huggingface.co/owensong/Inflect-Atom-v3/tree/main"><img src="assets/v3-link-download.svg" alt="DOWNLOAD" height="34"></a>
26
+ <a href="#voices"><img src="assets/v3-link-voices.svg" alt="SIX VOICES" height="34"></a>
27
+ <a href="#family"><img src="assets/v3-link-family.svg" alt="THE FAMILY" height="34"></a>
28
+ <a href="https://github.com/owenawsong/Inflect"><img src="assets/v3-link-github.svg" alt="GITHUB" height="34"></a>
29
+ <a href="https://discord.gg/CVJYedvzvp"><img src="assets/v3-link-discord.svg" alt="DISCORD" height="34"></a>
30
+ <a href="LICENSE.md"><img src="assets/v3-link-terms.svg" alt="TERMS" height="34"></a></p>
31
+
32
+ > **Usable early access, not an announced final launch.** This repository is a
33
+ > versioned package for noncommercial local evaluation. Feature availability is
34
+ > described below, not implied by a demo or a future roadmap. Existing Preview
35
+ > repositories remain separate and unchanged.
36
+
37
+ <p align="center"><img src="assets/v3-specs.svg" alt="157,190 synthesis-core parameters; six fixed English voices; native 24 kHz" width="100%"></p>
38
+
39
+ ## At a glance
40
+
41
+ | Item | Inflect Atom v3 |
42
+ | --- | --- |
43
+ | Learned synthesis core | **157,190 parameters** |
44
+ | Core weight file | `export.pt` |
45
+ | Serialized core weight file | **656,498 bytes** (0.66 MB, decimal) |
46
+ | Native output | 24 kHz English speech |
47
+ | Voice IDs | `hope`, `sophia`, `cassidy`, `james`, `jasper`, `mark` |
48
+ | Integrity inventory | [`manifest.json`](manifest.json) |
49
+
50
+ Core parameters, serialized weight size, full download size, installed dependencies
51
+ and runtime memory are different quantities. The core count does **not** include
52
+ external pronunciation assets or fixed voice-conditioning assets. The core weight
53
+ file is **not the complete installation**; frontend/runtime/vendor assets are also
54
+ required. Optional Mini components, where applicable, are counted separately.
55
+
56
+ ## Run locally
57
+
58
+ Download the complete repository, not only the weight file. Use your own Hugging
59
+ Face account if its current access controls require sign-in:
60
+
61
+ ```sh
62
+ python -m pip install huggingface-hub==0.36.2
63
+ # Only if authentication is required: hf auth login
64
+ hf download owensong/Inflect-Atom-v3 --local-dir inflect-atom-v3
65
+ cd inflect-atom-v3
66
+ ```
67
+
68
+ Use Python 3.12 and install the base dependencies plus the bundled pronunciation wheel:
69
+
70
+ ```sh
71
+ python -m pip install -r requirements.txt
72
+ python -m pip install vendor/en_core_web_sm-3.8.0-py3-none-any.whl
73
+ python synthesize.py --verify
74
+ python synthesize.py --voice cassidy --text "A small idea can become something worth hearing." --output ../atom-cassidy.wav
75
+ ```
76
+
77
+ The wrapper selects this package's model directory; there is no `--model atom`
78
+ flag. The default frontend is Misaki, the default waveform backend is `torch`,
79
+ and the default seed is 7717. Output is 24 kHz speech. Optional `cpu-fused` and
80
+ `cpu-lut` backends have additional dependencies in `requirements-cpu.txt`; that
81
+ file supplements the base requirements and is **not** a complete standalone
82
+ installation recipe. Alternate backends/frontends are not equivalence claims.
83
+
84
+ Choose a **new WAV path whose parent directory already exists**. The examples
85
+ write beside the downloaded package, keeping its integrity-checked files intact;
86
+ existing outputs are never overwritten. Once dependencies are installed and all
87
+ package assets are present, generation is local. Installation may need network
88
+ access; "local inference" is not a promise that setup downloads nothing.
89
+
90
+ These commands follow this package's CLI and dependency files. They are not a
91
+ blanket clean-install, GPU, platform or cross-runtime bitwise-compatibility claim.
92
+ Unsupported text fails explicitly rather than silently promising every script,
93
+ symbol or language. Check the returned metadata when testing numbers and chunks.
94
+
95
+ ## Voices
96
+
97
+ All six are **preset voice IDs**, not a reference-audio cloning interface. Change
98
+ the `--voice` argument to compare them on your own text; no external voice
99
+ recording is needed for these presets.
100
+
101
+ | Preset | CLI ID |
102
+ | --- | --- |
103
+ | Hope | `hope` |
104
+ | Sophia | `sophia` |
105
+ | Cassidy | `cassidy` |
106
+ | James | `james` |
107
+ | Jasper | `jasper` |
108
+ | Mark | `mark` |
109
+
110
+ No earlier Preview sample is relabeled as output from these weights. For a fair
111
+ comparison, generate the same text with the exact revision, voice, device, seed
112
+ and profile/backend you intend to use. Public synthetic examples should be
113
+ clearly identified as synthetic and respect the applicable terms.
114
+
115
+ ## Features and boundaries
116
+
117
+ | Capability | This download |
118
+ | --- | --- |
119
+ | Original six fixed English voices | Included, 24 kHz |
120
+ | Cassidy Q1500 / Mini's optional 48 kHz enhancer | Not included; Mini-specific components |
121
+ | Reference-based Tuner or voice cloning | Not exposed by this package |
122
+ | Generated event tags or learned style tags | Not exposed by this package |
123
+
124
+ An optional Mini component is not an upgrade to this model. Bracketed instructions
125
+ such as `[laugh]` or `[excited]` are **not supported feature controls** here.
126
+ The English voice presets are not an accent-conversion or cloning service.
127
+
128
+ ## Family
129
+
130
+ | v3 package | Learned synthesis core | Serialized core weight file, not total download |
131
+ | --- | ---: | ---: |
132
+ | [Inflect Mini v3](https://huggingface.co/owensong/Inflect-Mini-v3) | 34,310,790 | 138.39 MB |
133
+ | [Inflect Micro v3](https://huggingface.co/owensong/Inflect-Micro-v3) | 11,787,634 | 48.06 MB |
134
+ | [Inflect Atom v3](https://huggingface.co/owensong/Inflect-Atom-v3) | 157,190 | 0.66 MB |
135
+
136
+ Mini prioritizes the larger speech core, Micro offers a smaller neural option,
137
+ and Atom explores an exceptionally small synthesis core without a separate
138
+ learned neural vocoder. Choose by listening to your use case, not parameter count
139
+ alone. These new repositories are not aliases for the historical Preview
140
+ downloads, and historical Micro Preview results do not automatically describe
141
+ the different FastMicro core here. Nano is not part of this three-package release.
142
+
143
+ ## Evidence, terms and notices
144
+
145
+ No official leaderboard position, human MOS, full-clone capability or general
146
+ quality improvement is asserted by this card. A known regression/development
147
+ panel is not an unseen test. Source-recording ASR, model-output WER, acoustic
148
+ proxy scores and human listening are different evidence and must not be conflated.
149
+ No new WER/MOS figure is assigned to an optional or unqualified route here.
150
+
151
+ **Noncommercial evaluation only.** The exact
152
+ [Inflect Preview Evaluation License 1.0](LICENSE.md) governs the identified
153
+ Inflect-specific assets, only to the extent of rights held or authorized by the
154
+ licensor. Commercial use, redistribution and paid **or free** hosted services
155
+ require separate written permission under those terms. This is not an open-source
156
+ model license or a grant of third-party voice/identity rights.
157
+
158
+ Read [component scope](LICENSE_SCOPE.json), [third-party notices](THIRD_PARTY_NOTICES.md)
159
+ and [access status](BETA_ACCESS.md). Applicable upstream terms and earlier
160
+ independently valid grants remain intact; in particular, retain the exact supplied
161
+ [Apache-2.0](LICENSES/Apache-2.0.txt) and [StyleTTS2 MIT](LICENSES/StyleTTS2-MIT.txt)
162
+ texts where included. Nothing here relicenses upstream work. Dependency and
163
+ bundled pronunciation-wheel licenses remain their own. Training recipes, private
164
+ source recordings and development caches are not distributed by this card.
165
+
166
+ Questions or reproducible issues: [repository discussions](https://huggingface.co/owensong/Inflect-Atom-v3/discussions).
167
+ Project: [GitHub](https://github.com/owenawsong/Inflect) / [Discord](https://discord.gg/CVJYedvzvp). Include the revision, voice,
168
+ profile/backend and a minimal text example; never post credentials or private recordings.
assets/banner-v3.svg ADDED
assets/v3-link-discord.svg ADDED
assets/v3-link-download.svg ADDED
assets/v3-link-family.svg ADDED
assets/v3-link-github.svg ADDED
assets/v3-link-run.svg ADDED
assets/v3-link-terms.svg ADDED
assets/v3-link-voices.svg ADDED
assets/v3-specs.svg ADDED
manifest.json CHANGED
@@ -22,17 +22,53 @@
22
  "size": 312
23
  },
24
  "README.md": {
25
- "sha256": "3059570f55ad1359065f38318b2f94105c04a9e3d8b1808dfdb18dd25196ecde",
26
- "size": 944
27
  },
28
  "THIRD_PARTY_NOTICES.md": {
29
  "sha256": "8c939b9eb4538c7fb0dcfc72916f75d56e025ba4c75fe3f838a5f1efa472080e",
30
  "size": 1262
31
  },
 
 
 
 
32
  "assets/inflect-mark.svg": {
33
  "sha256": "49a442da2b2130010f39778e58abdce7154a17cb1920dc4bcb5d0b8a3818156b",
34
  "size": 301
35
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
36
  "atom_dsp_cpu_20261024.py": {
37
  "sha256": "07df5e1596133310a6a24e2a3e2269985942241acbdbae93778348ff4e88494b",
38
  "size": 8331
 
22
  "size": 312
23
  },
24
  "README.md": {
25
+ "sha256": "74009c8cad261cbb0e1c417184489a5626692c7ba6aef8e6349870a146ddef05",
26
+ "size": 8504
27
  },
28
  "THIRD_PARTY_NOTICES.md": {
29
  "sha256": "8c939b9eb4538c7fb0dcfc72916f75d56e025ba4c75fe3f838a5f1efa472080e",
30
  "size": 1262
31
  },
32
+ "assets/banner-v3.svg": {
33
+ "sha256": "5f13e67ef56269521f27f893097a486bbecb73cfea7c3bc98982b3219946b483",
34
+ "size": 2280
35
+ },
36
  "assets/inflect-mark.svg": {
37
  "sha256": "49a442da2b2130010f39778e58abdce7154a17cb1920dc4bcb5d0b8a3818156b",
38
  "size": 301
39
  },
40
+ "assets/v3-link-discord.svg": {
41
+ "sha256": "a5e210ce3699ae16902d6216795d820871c5187738e0de9ffdfe9e5706c49083",
42
+ "size": 404
43
+ },
44
+ "assets/v3-link-download.svg": {
45
+ "sha256": "eda2fb31f287eaf45e31431ce05d07d16ed32c1f5bbcf70e8e8cfd3d2147d12b",
46
+ "size": 406
47
+ },
48
+ "assets/v3-link-family.svg": {
49
+ "sha256": "c0e31f3008d5989fb784d02373b3cb593507497b6b3414c7911ac0555cb997dc",
50
+ "size": 413
51
+ },
52
+ "assets/v3-link-github.svg": {
53
+ "sha256": "c021e227917583970fcd2d44bc54a61fa5e2880637288c7580db9683a2d17f7f",
54
+ "size": 402
55
+ },
56
+ "assets/v3-link-run.svg": {
57
+ "sha256": "3a0928fbb5b0f9f33300ffc1f86f4754dfeca5112ac980a6a21a75be1eb729d7",
58
+ "size": 415
59
+ },
60
+ "assets/v3-link-terms.svg": {
61
+ "sha256": "f178b2ea4088b859c66a3759b8e8849d5c3813cd853bc4fbdf75d5617c0ca0b0",
62
+ "size": 400
63
+ },
64
+ "assets/v3-link-voices.svg": {
65
+ "sha256": "54344875395489d95124ed86ff58e6e06ab42627a8b57f4bda8c1e1891f831d7",
66
+ "size": 413
67
+ },
68
+ "assets/v3-specs.svg": {
69
+ "sha256": "8bebe02e92ecab6a0cba8a0b4844706fc0de609712f0f591b66fed2c05c2ca25",
70
+ "size": 711
71
+ },
72
  "atom_dsp_cpu_20261024.py": {
73
  "sha256": "07df5e1596133310a6a24e2a3e2269985942241acbdbae93778348ff4e88494b",
74
  "size": 8331