owensong commited on
Commit
e2407ee
·
verified ·
1 Parent(s): 94ea621

Use absolute Hub audio URLs for model-card playback

Browse files
Files changed (2) hide show
  1. README.md +294 -294
  2. release-manifest.json +2 -2
README.md CHANGED
@@ -1,294 +1,294 @@
1
- ---
2
- language:
3
- - en
4
- pipeline_tag: text-to-speech
5
- model_name: Inflect Mini 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-Mini-v3/resolve/main/assets/banner-v3.svg
17
- ---
18
-
19
- <p align="center"><img src="assets/banner-v3.svg" alt="Inflect Mini v3. Six voices. A 34.3M-parameter speech core. Local generation, with optional 48 kHz enhancement." width="100%"></p>
20
-
21
- <h1 align="center">A small model. An open-ended canvas.</h1>
22
-
23
- <p align="center"><strong>Six voices for the things you want to say.</strong><br>
24
- English text to speech, running locally on your own hardware.<br>
25
- Start with the core. Add detail, delivery and event controls when you need them.</p>
26
-
27
- <p align="center">
28
- <a href="#hear-mini"><img src="assets/mini-link-listen.svg" alt="LISTEN" height="32"></a>
29
- <a href="#run-locally"><img src="assets/mini-link-run.svg" alt="RUN LOCALLY" height="32"></a>
30
- <a href="#evaluations"><img src="assets/mini-link-evidence.svg" alt="EVALUATIONS" height="32"></a>
31
- <a href="https://github.com/owenawsong/Inflect"><img src="assets/mini-link-github.svg" alt="GITHUB" height="32"></a>
32
- <a href="https://discord.gg/CVJYedvzvp"><img src="assets/mini-link-discord.svg" alt="DISCORD" height="32"></a>
33
- </p>
34
-
35
- > **Early access.** This is a usable, noncommercial evaluation release, not the announced final launch. The features below describe what is available today. [Read the terms](LICENSE.md).
36
-
37
- <p align="center"><img src="assets/mini-specification.svg" alt="34.3M core parameters; six preset English voices; native 24 kHz; optional 48 kHz enhancement." width="100%"></p>
38
-
39
- ## Hear Mini
40
-
41
- Before the specifications, the speech. These samples come from **Original Mini at native 24 kHz**, with no enhancer or teacher audio at inference. Each voice reads a different published input; they are examples, not a controlled voice-ranking test.
42
-
43
- ### Cassidy
44
-
45
- > Could you bring the smaller suitcase downstairs?
46
-
47
- <audio controls preload="none" src="media/mini-cassidy.wav"></audio>
48
-
49
- ### Mark
50
-
51
- > Beyond the orchard, a narrow footpath curved toward the ruined tower.
52
-
53
- <audio controls preload="none" src="media/mini-mark.wav"></audio>
54
-
55
- <details>
56
- <summary><strong>Meet the other four voices</strong></summary>
57
-
58
- ### Hope
59
-
60
- > Press the side button once, then wait until the indicator turns green.
61
-
62
- <audio controls preload="none" src="media/mini-hope.wav"></audio>
63
-
64
- ### James
65
-
66
- > A shadow becomes longer when the light source moves closer to the horizon.
67
-
68
- <audio controls preload="none" src="media/mini-james.wav"></audio>
69
-
70
- ### Sophia
71
-
72
- > We moved the chairs inside and wiped the wooden table dry.
73
-
74
- <audio controls preload="none" src="media/mini-sophia.wav"></audio>
75
-
76
- ### Jasper
77
-
78
- > I thought you were arriving tomorrow, but it is lovely to see you.
79
-
80
- <audio controls preload="none" src="media/mini-jasper.wav"></audio>
81
-
82
- </details>
83
-
84
- ## One core. Choose your finish.
85
-
86
- Mini is designed to be useful as a straightforward local speech model. Its optional components are explicit choices, not hidden requirements for ordinary generation.
87
-
88
- | Start here | Add when you need it |
89
- | :--- | :--- |
90
- | **Six preset voices**<br>Hope, Sophia, Cassidy, James, Jasper and Mark. Choose a voice and write your text. | **48 kHz detail**<br>A separate 2.85M-parameter enhancer extends the native 24 kHz output to a 48 kHz WAV. |
91
- | **Local generation**<br>Speech runs on your machine after setup. CPU is the default; compatible CUDA installations are supported. | **Cassidy delivery**<br>The optional Q1500 pitch profile changes Cassidy's delivery while retaining native timing. |
92
- | **Plain-text speech**<br>The base path stays available without event or adaptation dependencies. | **Event beta and Tuner SDK**<br>Four selected Cassidy event performances, plus a separate experimental reference-voice adaptation workflow. |
93
-
94
- ### A little more detail, at 48 kHz
95
-
96
- Hear the same Mini generation before and after the optional enhancer. The words and underlying speech are the same; this is bandwidth extension, not a different voice or a native-48 kHz speech backbone.
97
-
98
- > Could you bring the smaller suitcase downstairs?
99
-
100
- | Original Mini / 24 kHz | Mini + enhancer / 48 kHz |
101
- | :--- | :--- |
102
- | <audio controls preload="none" src="media/mini-cassidy.wav"></audio> | <audio controls preload="none" src="media/mini-cassidy-48k.wav"></audio> |
103
-
104
- Both UTMOS readers used in the reference evaluation operate internally at 16 kHz. They do not directly measure the added detail above 8 kHz; listening remains important.
105
-
106
- ### Give a line a moment
107
-
108
- ```text
109
- You really surprised me. [laugh] That was a good one.
110
- ```
111
-
112
- The **Cassidy event beta** supports `[laugh]`, `[sigh]`, `[gasp]` and `[cough]` through a separate event route. The selected takes passed human listening review. Speech is generated around the event, with the joins handled by the event component.
113
-
114
- **Scope matters:** each tag reuses one fixed learned performance. These are not newly varied events on every request, emotion controls, or support for every preset voice. The published event command below produces 24 kHz audio; it does not apply the speech enhancer. [Event guide](extras/events/README.md).
115
-
116
- ### Explore another voice, experimentally
117
-
118
- The optional **Tuner SDK** builds a saved adaptation profile from an authorized 5-15 second reference. Subsequent playback uses Mini and the saved profile without needing the recording or enrollment encoders during speech generation.
119
-
120
- This adapts Mini's existing voice conditioning. **It is not accurate, unrestricted voice cloning**, and identity, accent and quality are not guaranteed. It is optional, with its own setup, encoder downloads and license requirements. [Tuner guide](extras/tuner/README.md).
121
-
122
- ## Run locally
123
-
124
- A complete download, Python 3.12, and one command to speak. Download the **whole repository**, not just the core weight file.
125
-
126
- ```sh
127
- python -m pip install huggingface-hub==0.36.2
128
- hf download owensong/Inflect-Mini-v3 --local-dir inflect-mini-v3
129
- cd inflect-mini-v3
130
- python verify_release.py
131
- python -m pip install -r core/requirements-cpu.txt
132
- python run.py speech --voice cassidy --text "A small idea can become something worth hearing." --output ../mini-cassidy.wav
133
- ```
134
-
135
- Use `hope`, `sophia`, `cassidy`, `james`, `jasper` or `mark`. The default is Original Mini, CPU, canonical runtime, native 24 kHz and seed 42. Choose a new output file outside the package, in an existing directory.
136
-
137
- <details>
138
- <summary><strong>Choose delivery or 48 kHz output</strong></summary>
139
-
140
- ```sh
141
- # Optional Cassidy delivery, native timing retained.
142
- python run.py speech --profile pitch --voice cassidy --text "Are we ready to begin?" --output ../mini-delivery.wav
143
-
144
- # Optional 48 kHz enhancement, available for all six speech voices.
145
- python run.py speech --profile enhanced --voice hope --text "The next chapter starts here." --output ../mini-48k.wav
146
-
147
- # Cassidy delivery plus 48 kHz enhancement.
148
- python run.py speech --profile combined --voice cassidy --text "Are we ready to begin?" --output ../mini-delivery-48k.wav
149
- ```
150
-
151
- The pitch option is **Cassidy Q1500 only**; other voices keep Original delivery. A physical guard can revert an affected chunk to Original, and the delivery metadata records the route taken. The newer joint-polish and emotion experiments are not included.
152
-
153
- </details>
154
-
155
- <details>
156
- <summary><strong>Use the four Cassidy event tags</strong></summary>
157
-
158
- ```sh
159
- python -m pip install -r extras/events/component/requirements-cpu.txt
160
- python run.py events --voice cassidy --profile base --text "You really surprised me. [laugh] That was a good one." --output ../mini-laugh.wav
161
- ```
162
-
163
- Replace `[laugh]` with `[sigh]`, `[gasp]` or `[cough]`. Events are off by default and require the event launcher, not the ordinary speech command. `--profile pitch` selects Cassidy Q1500. `--laugh-edge-room-ms 20` optionally adds 20 ms at each adjoining laugh boundary without cutting the selected event or breath samples.
164
-
165
- </details>
166
-
167
- <details>
168
- <summary><strong>CUDA, output format and offline operation</strong></summary>
169
-
170
- For CUDA, use a compatible installation following `core/requirements-cuda.txt`, then add `--device cuda`. The CPU recipe above is not a CUDA installation recipe. FLOAT WAV is the speech default; `--subtype PCM_16` selects PCM16.
171
-
172
- Speech generation is local once the complete model, pronunciation assets and dependencies are installed. Initial setup can require downloads. Optional Tuner enrollment has separate encoder-download requirements.
173
-
174
- The public `run.py` verifies the release inventory and disables model bytecode writes before launching the selected CLI. Keep generated outputs outside the integrity-checked package.
175
-
176
- </details>
177
-
178
- <details>
179
- <summary><strong>Enroll an authorized reference with the optional Tuner</strong></summary>
180
-
181
- The tested enrollment setup is Linux CPU, using Python 3.12. Install its optional dependencies in your active environment:
182
-
183
- ```sh
184
- python -m pip install uv
185
- python -m uv pip install --python python --find-links core/baseline/vendor -c extras/tuner/constraints-cpu.txt -r extras/tuner/requirements-tuner-enrollment-cpu-20261110.txt
186
- ```
187
-
188
- Place an authorized recording at `../my-reference.wav`. The profile and work directory must be new, outside the download, and their parent directories must exist.
189
-
190
- ```sh
191
- python tuner.py enroll --reference ../my-reference.wav --profile ../my-voice.json --work-dir ../tuner-work --device cpu --consent-confirmed --download-public-assets --accept-third-party-licenses
192
- python tuner.py say --profile ../my-voice.json --work-dir ../tuner-work --device cpu --text "A new chapter starts here." --output ../my-voice.wav
193
- ```
194
-
195
- Keep the profile JSON, its companion `.pt` file and work directory. Use only voices and recordings you are authorized to enroll. The SDK contains no reference recordings, user profiles or enrollment encoder weights. See the [advanced setup](extras/tuner/README.md) and [component license map](extras/tuner/LICENSE-MAP.json).
196
-
197
- </details>
198
-
199
- ## Evaluations
200
-
201
- **Clear scope is part of a useful result.** This section separates the measurements we have from the evaluations still needed for this exact published configuration. A dash is a placeholder, not a score of zero.
202
-
203
- ### Release-matched evaluation board
204
-
205
- | Evaluation | Result | Status |
206
- | :--- | :---: | :--- |
207
- | Open TTS Leaderboard / average English WER | **&mdash;** | Not run and verified for this exact published configuration. |
208
- | Open TTS Leaderboard / speaker similarity | **&mdash;** | Not run for the published Tuner; no comparable cloning claim. |
209
- | UTMOS22 / complete current release report | **&mdash;** | Not run as a release-matched report; historical reference panel below. |
210
- | UTMOS v2 / complete current release report | **&mdash;** | Not run as a release-matched report; historical reference panel below. |
211
- | TTSDS / release-matched distribution evaluation | **&mdash;** | Not run for this published configuration. |
212
- | Human preference / blinded release study | **&mdash;** | Not run. Listening review is not a controlled preference study. |
213
-
214
- These placeholders are intentionally visible. They must not be used to infer a leaderboard position, superiority to another model, or validation of optional features. Existing research diagnostics are not automatically scores for the released package.
215
-
216
- <details>
217
- <summary><strong>Existing reference measurements: what they do and do not establish</strong></summary>
218
-
219
- The November 6 selected-core replay evaluated Original Mini and the retained enhancer with the same underlying core weights used here. **This is historical reference evidence, not a fresh end-to-end benchmark of this public package.** The source/frontends, package wrapper and optional feature scope must remain attached to the result.
220
-
221
- | Selected-core reference | Original Mini | Mini + enhancer |
222
- | :--- | ---: | ---: |
223
- | WER / 144 clips, 2,011 reference words | 1.591% | 1.591% |
224
- | Recognition edits | 32 | 32 |
225
- | UTMOS22 Strong / 24 voice-text pairs | 4.4600 | 4.4599 |
226
- | UTMOS v2 fold0 / same 24 pairs | 3.9234 | 3.9261 |
227
- | Warm CPU throughput / times real-time | 3.97x | 3.69x |
228
- | Warm RTX 5090 throughput / times real-time | 55.17x | 53.45x |
229
-
230
- **Protocol.** Deterministic replay balanced across all six preset voices, not a certified training-disjoint holdout or the official Open TTS leaderboard. ASR used Whisper large-v3 CTranslate2, FP16, beam 5, no VAD or reference prompt. Both sides used the same documented text normalization; endings and spelling variants were not forgiven. The MOS readers are automatic quality proxies, not human opinion scores.
231
-
232
- **Timing.** Warm text preparation through returned waveform; 12 texts across six voices, three measured calls per text. CPU: Ryzen 7 7800X3D, two threads. GPU: RTX 5090. Loading, queueing, encoding and disk writes were excluded. These are not cold-start latencies or competitor comparisons. Higher times real-time is faster.
233
-
234
- **Scope.** These figures do not evaluate Q1500, the combined profile, event tags or Tuner quality. The enhancer differences on the two quality readers did not establish a clear aggregate gain or loss. Both readers internally use 16 kHz, so they do not directly score bandwidth above 8 kHz.
235
-
236
- Machine-readable provenance for the featured audio and reference numbers: [presentation evidence](assets/presentation-evidence.json).
237
-
238
- </details>
239
-
240
- ### Download checks, not just demo checks
241
-
242
- The November 10 anonymous CPU check downloaded the public Mini repository without a token, verified all **445 inventoried files**, and rendered ordinary speech plus each of the four event tags. The returned audio matched the corresponding qualified outputs exactly. This establishes a functional check on that published revision, not a new perceptual-quality score or universal platform certification.
243
-
244
- ## Small core. Explicit extras.
245
-
246
- <p align="center"><img src="assets/mini-components.svg" alt="Original Mini: 34.31M core parameters. Add 3.80M for Cassidy delivery, 2.85M for enhancement, or both for 40.96M total. Event and Tuner components are separate." width="100%"></p>
247
-
248
- | Speech configuration | Added learned parameters | Total learned speech parameters |
249
- | :--- | ---: | ---: |
250
- | Original Mini | 0 | **34,310,790** |
251
- | Mini + Cassidy Q1500 | 3,801,977 | **38,112,767** |
252
- | Mini + 48 kHz enhancer | 2,846,722 | **37,157,512** |
253
- | Cassidy Q1500 + enhancer | 6,648,699 | **40,959,489** |
254
-
255
- The base weight file is **138,388,547 bytes** (138.39 MB, decimal), at `core/baseline/models/mini/weights.pt`. Core parameters, weight-file size, complete download, installed dependencies and runtime memory are different quantities. Pronunciation assets and fixed voice-conditioning assets are not included in core parameter counts.
256
-
257
- The event beta separately adds a **13,531,650-parameter Vocos decoder** and **25,600 learned control values**. Tuner enrollment uses separately acquired encoders; their footprint is not included in Mini's core count. See each component's guide for setup and terms.
258
-
259
- ## A family, not one size for every job
260
-
261
- | | Mini v3 | Micro v3 | Atom v3 |
262
- | :--- | :--- | :--- | :--- |
263
- | Core parameters | **34.31M** | **11.79M** | **157.19K** |
264
- | Place in the family | Speech core with optional feature components | Smaller neural speech core | Ultra-compact synthesis core |
265
- | Preset English voices | 6 | 6 | 6 |
266
- | Native output | 24 kHz | 24 kHz | 24 kHz |
267
- | Get the model | **You are here** | [Micro](https://huggingface.co/owensong/Inflect-Micro-v3) | [Atom](https://huggingface.co/owensong/Inflect-Atom-v3) |
268
-
269
- Micro and Atom are separate models, not Mini with its optional components removed. Their capabilities and tradeoffs are documented on their own cards. Historical Preview repositories remain available and unchanged.
270
-
271
- ## Before you build
272
-
273
- Mini is a six-preset **English** model. Difficult pronunciation, numerals, long passages and delivery can still fail; review generated speech before using it. The 48 kHz option is enhancement, not a native-48 kHz backbone. Q1500 is Cassidy-only, and the four event performances are fixed. **Excited, angry, whisper and other research style tags are not supported in this release.** The experimental General frontend is not the default.
274
-
275
- Use this release for **noncommercial evaluation** under the exact [Inflect Preview Evaluation License 1.0](LICENSE.md). Commercial use, redistribution and paid or free hosted services require separate written permission under those terms. These terms do not override upstream licenses or grant third-party voice or identity rights.
276
-
277
- <details>
278
- <summary><strong>Package integrity, licenses and acknowledgments</strong></summary>
279
-
280
- `core/` contains the speech package; `extras/events/` and `extras/tuner/` contain independent optional components. The [outer release inventory](release-manifest.json), [Core inventory](core/assembly-manifest.json) and nested component manifests record separate compatibility identities. This presentation update does not change speech weights or runtime code.
281
-
282
- Read the [top-level notices](THIRD_PARTY_NOTICES.md), [Core license scope](core/LICENSE_SCOPE.json), [Core notices](core/THIRD_PARTY_NOTICES.md), and the notices supplied with each optional component. Applicable [StyleTTS2 MIT](core/LICENSES/StyleTTS2-MIT.txt), [Apache-2.0](core/LICENSES/Apache-2.0.txt) and [NVIDIA REUSE OpenMDW-1.1](core/LICENSES/NVIDIA_REUSE_OpenMDW-1.1.txt) texts are retained. The REUSE teacher is not bundled. The event decoder has its own MIT notice. Tuner encoder terms include Apache-2.0 and an embedded CC BY-SA 3.0 speaker encoder; their obligations remain separate from Inflect's terms.
283
-
284
- Private recordings and development recipes are not part of the public package. Required attribution is preserved without presenting a training recipe.
285
-
286
- </details>
287
-
288
- ---
289
-
290
- <p align="center"><strong>Make something worth hearing.</strong><br>
291
- <a href="#run-locally">Run Mini locally</a> &nbsp; / &nbsp;
292
- <a href="https://github.com/owenawsong/Inflect">Explore Inflect</a> &nbsp; / &nbsp;
293
- <a href="https://discord.gg/CVJYedvzvp">Join the conversation</a> &nbsp; / &nbsp;
294
- <a href="https://huggingface.co/owensong/Inflect-Mini-v3/discussions">Feedback and issues</a></p>
 
1
+ ---
2
+ language:
3
+ - en
4
+ pipeline_tag: text-to-speech
5
+ model_name: Inflect Mini 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-Mini-v3/resolve/main/assets/banner-v3.svg
17
+ ---
18
+
19
+ <p align="center"><img src="assets/banner-v3.svg" alt="Inflect Mini v3. Six voices. A 34.3M-parameter speech core. Local generation, with optional 48 kHz enhancement." width="100%"></p>
20
+
21
+ <h1 align="center">A small model. An open-ended canvas.</h1>
22
+
23
+ <p align="center"><strong>Six voices for the things you want to say.</strong><br>
24
+ English text to speech, running locally on your own hardware.<br>
25
+ Start with the core. Add detail, delivery and event controls when you need them.</p>
26
+
27
+ <p align="center">
28
+ <a href="#hear-mini"><img src="assets/mini-link-listen.svg" alt="LISTEN" height="32"></a>
29
+ <a href="#run-locally"><img src="assets/mini-link-run.svg" alt="RUN LOCALLY" height="32"></a>
30
+ <a href="#evaluations"><img src="assets/mini-link-evidence.svg" alt="EVALUATIONS" height="32"></a>
31
+ <a href="https://github.com/owenawsong/Inflect"><img src="assets/mini-link-github.svg" alt="GITHUB" height="32"></a>
32
+ <a href="https://discord.gg/CVJYedvzvp"><img src="assets/mini-link-discord.svg" alt="DISCORD" height="32"></a>
33
+ </p>
34
+
35
+ > **Early access.** This is a usable, noncommercial evaluation release, not the announced final launch. The features below describe what is available today. [Read the terms](LICENSE.md).
36
+
37
+ <p align="center"><img src="assets/mini-specification.svg" alt="34.3M core parameters; six preset English voices; native 24 kHz; optional 48 kHz enhancement." width="100%"></p>
38
+
39
+ ## Hear Mini
40
+
41
+ Before the specifications, the speech. These samples come from **Original Mini at native 24 kHz**, with no enhancer or teacher audio at inference. Each voice reads a different published input; they are examples, not a controlled voice-ranking test.
42
+
43
+ ### Cassidy
44
+
45
+ > Could you bring the smaller suitcase downstairs?
46
+
47
+ <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-cassidy.wav"></audio>
48
+
49
+ ### Mark
50
+
51
+ > Beyond the orchard, a narrow footpath curved toward the ruined tower.
52
+
53
+ <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-mark.wav"></audio>
54
+
55
+ <details>
56
+ <summary><strong>Meet the other four voices</strong></summary>
57
+
58
+ ### Hope
59
+
60
+ > Press the side button once, then wait until the indicator turns green.
61
+
62
+ <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-hope.wav"></audio>
63
+
64
+ ### James
65
+
66
+ > A shadow becomes longer when the light source moves closer to the horizon.
67
+
68
+ <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-james.wav"></audio>
69
+
70
+ ### Sophia
71
+
72
+ > We moved the chairs inside and wiped the wooden table dry.
73
+
74
+ <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-sophia.wav"></audio>
75
+
76
+ ### Jasper
77
+
78
+ > I thought you were arriving tomorrow, but it is lovely to see you.
79
+
80
+ <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-jasper.wav"></audio>
81
+
82
+ </details>
83
+
84
+ ## One core. Choose your finish.
85
+
86
+ Mini is designed to be useful as a straightforward local speech model. Its optional components are explicit choices, not hidden requirements for ordinary generation.
87
+
88
+ | Start here | Add when you need it |
89
+ | :--- | :--- |
90
+ | **Six preset voices**<br>Hope, Sophia, Cassidy, James, Jasper and Mark. Choose a voice and write your text. | **48 kHz detail**<br>A separate 2.85M-parameter enhancer extends the native 24 kHz output to a 48 kHz WAV. |
91
+ | **Local generation**<br>Speech runs on your machine after setup. CPU is the default; compatible CUDA installations are supported. | **Cassidy delivery**<br>The optional Q1500 pitch profile changes Cassidy's delivery while retaining native timing. |
92
+ | **Plain-text speech**<br>The base path stays available without event or adaptation dependencies. | **Event beta and Tuner SDK**<br>Four selected Cassidy event performances, plus a separate experimental reference-voice adaptation workflow. |
93
+
94
+ ### A little more detail, at 48 kHz
95
+
96
+ Hear the same Mini generation before and after the optional enhancer. The words and underlying speech are the same; this is bandwidth extension, not a different voice or a native-48 kHz speech backbone.
97
+
98
+ > Could you bring the smaller suitcase downstairs?
99
+
100
+ | Original Mini / 24 kHz | Mini + enhancer / 48 kHz |
101
+ | :--- | :--- |
102
+ | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-cassidy.wav"></audio> | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Mini-v3/resolve/main/media/mini-cassidy-48k.wav"></audio> |
103
+
104
+ Both UTMOS readers used in the reference evaluation operate internally at 16 kHz. They do not directly measure the added detail above 8 kHz; listening remains important.
105
+
106
+ ### Give a line a moment
107
+
108
+ ```text
109
+ You really surprised me. [laugh] That was a good one.
110
+ ```
111
+
112
+ The **Cassidy event beta** supports `[laugh]`, `[sigh]`, `[gasp]` and `[cough]` through a separate event route. The selected takes passed human listening review. Speech is generated around the event, with the joins handled by the event component.
113
+
114
+ **Scope matters:** each tag reuses one fixed learned performance. These are not newly varied events on every request, emotion controls, or support for every preset voice. The published event command below produces 24 kHz audio; it does not apply the speech enhancer. [Event guide](extras/events/README.md).
115
+
116
+ ### Explore another voice, experimentally
117
+
118
+ The optional **Tuner SDK** builds a saved adaptation profile from an authorized 5-15 second reference. Subsequent playback uses Mini and the saved profile without needing the recording or enrollment encoders during speech generation.
119
+
120
+ This adapts Mini's existing voice conditioning. **It is not accurate, unrestricted voice cloning**, and identity, accent and quality are not guaranteed. It is optional, with its own setup, encoder downloads and license requirements. [Tuner guide](extras/tuner/README.md).
121
+
122
+ ## Run locally
123
+
124
+ A complete download, Python 3.12, and one command to speak. Download the **whole repository**, not just the core weight file.
125
+
126
+ ```sh
127
+ python -m pip install huggingface-hub==0.36.2
128
+ hf download owensong/Inflect-Mini-v3 --local-dir inflect-mini-v3
129
+ cd inflect-mini-v3
130
+ python verify_release.py
131
+ python -m pip install -r core/requirements-cpu.txt
132
+ python run.py speech --voice cassidy --text "A small idea can become something worth hearing." --output ../mini-cassidy.wav
133
+ ```
134
+
135
+ Use `hope`, `sophia`, `cassidy`, `james`, `jasper` or `mark`. The default is Original Mini, CPU, canonical runtime, native 24 kHz and seed 42. Choose a new output file outside the package, in an existing directory.
136
+
137
+ <details>
138
+ <summary><strong>Choose delivery or 48 kHz output</strong></summary>
139
+
140
+ ```sh
141
+ # Optional Cassidy delivery, native timing retained.
142
+ python run.py speech --profile pitch --voice cassidy --text "Are we ready to begin?" --output ../mini-delivery.wav
143
+
144
+ # Optional 48 kHz enhancement, available for all six speech voices.
145
+ python run.py speech --profile enhanced --voice hope --text "The next chapter starts here." --output ../mini-48k.wav
146
+
147
+ # Cassidy delivery plus 48 kHz enhancement.
148
+ python run.py speech --profile combined --voice cassidy --text "Are we ready to begin?" --output ../mini-delivery-48k.wav
149
+ ```
150
+
151
+ The pitch option is **Cassidy Q1500 only**; other voices keep Original delivery. A physical guard can revert an affected chunk to Original, and the delivery metadata records the route taken. The newer joint-polish and emotion experiments are not included.
152
+
153
+ </details>
154
+
155
+ <details>
156
+ <summary><strong>Use the four Cassidy event tags</strong></summary>
157
+
158
+ ```sh
159
+ python -m pip install -r extras/events/component/requirements-cpu.txt
160
+ python run.py events --voice cassidy --profile base --text "You really surprised me. [laugh] That was a good one." --output ../mini-laugh.wav
161
+ ```
162
+
163
+ Replace `[laugh]` with `[sigh]`, `[gasp]` or `[cough]`. Events are off by default and require the event launcher, not the ordinary speech command. `--profile pitch` selects Cassidy Q1500. `--laugh-edge-room-ms 20` optionally adds 20 ms at each adjoining laugh boundary without cutting the selected event or breath samples.
164
+
165
+ </details>
166
+
167
+ <details>
168
+ <summary><strong>CUDA, output format and offline operation</strong></summary>
169
+
170
+ For CUDA, use a compatible installation following `core/requirements-cuda.txt`, then add `--device cuda`. The CPU recipe above is not a CUDA installation recipe. FLOAT WAV is the speech default; `--subtype PCM_16` selects PCM16.
171
+
172
+ Speech generation is local once the complete model, pronunciation assets and dependencies are installed. Initial setup can require downloads. Optional Tuner enrollment has separate encoder-download requirements.
173
+
174
+ The public `run.py` verifies the release inventory and disables model bytecode writes before launching the selected CLI. Keep generated outputs outside the integrity-checked package.
175
+
176
+ </details>
177
+
178
+ <details>
179
+ <summary><strong>Enroll an authorized reference with the optional Tuner</strong></summary>
180
+
181
+ The tested enrollment setup is Linux CPU, using Python 3.12. Install its optional dependencies in your active environment:
182
+
183
+ ```sh
184
+ python -m pip install uv
185
+ python -m uv pip install --python python --find-links core/baseline/vendor -c extras/tuner/constraints-cpu.txt -r extras/tuner/requirements-tuner-enrollment-cpu-20261110.txt
186
+ ```
187
+
188
+ Place an authorized recording at `../my-reference.wav`. The profile and work directory must be new, outside the download, and their parent directories must exist.
189
+
190
+ ```sh
191
+ python tuner.py enroll --reference ../my-reference.wav --profile ../my-voice.json --work-dir ../tuner-work --device cpu --consent-confirmed --download-public-assets --accept-third-party-licenses
192
+ python tuner.py say --profile ../my-voice.json --work-dir ../tuner-work --device cpu --text "A new chapter starts here." --output ../my-voice.wav
193
+ ```
194
+
195
+ Keep the profile JSON, its companion `.pt` file and work directory. Use only voices and recordings you are authorized to enroll. The SDK contains no reference recordings, user profiles or enrollment encoder weights. See the [advanced setup](extras/tuner/README.md) and [component license map](extras/tuner/LICENSE-MAP.json).
196
+
197
+ </details>
198
+
199
+ ## Evaluations
200
+
201
+ **Clear scope is part of a useful result.** This section separates the measurements we have from the evaluations still needed for this exact published configuration. A dash is a placeholder, not a score of zero.
202
+
203
+ ### Release-matched evaluation board
204
+
205
+ | Evaluation | Result | Status |
206
+ | :--- | :---: | :--- |
207
+ | Open TTS Leaderboard / average English WER | **&mdash;** | Not run and verified for this exact published configuration. |
208
+ | Open TTS Leaderboard / speaker similarity | **&mdash;** | Not run for the published Tuner; no comparable cloning claim. |
209
+ | UTMOS22 / complete current release report | **&mdash;** | Not run as a release-matched report; historical reference panel below. |
210
+ | UTMOS v2 / complete current release report | **&mdash;** | Not run as a release-matched report; historical reference panel below. |
211
+ | TTSDS / release-matched distribution evaluation | **&mdash;** | Not run for this published configuration. |
212
+ | Human preference / blinded release study | **&mdash;** | Not run. Listening review is not a controlled preference study. |
213
+
214
+ These placeholders are intentionally visible. They must not be used to infer a leaderboard position, superiority to another model, or validation of optional features. Existing research diagnostics are not automatically scores for the released package.
215
+
216
+ <details>
217
+ <summary><strong>Existing reference measurements: what they do and do not establish</strong></summary>
218
+
219
+ The November 6 selected-core replay evaluated Original Mini and the retained enhancer with the same underlying core weights used here. **This is historical reference evidence, not a fresh end-to-end benchmark of this public package.** The source/frontends, package wrapper and optional feature scope must remain attached to the result.
220
+
221
+ | Selected-core reference | Original Mini | Mini + enhancer |
222
+ | :--- | ---: | ---: |
223
+ | WER / 144 clips, 2,011 reference words | 1.591% | 1.591% |
224
+ | Recognition edits | 32 | 32 |
225
+ | UTMOS22 Strong / 24 voice-text pairs | 4.4600 | 4.4599 |
226
+ | UTMOS v2 fold0 / same 24 pairs | 3.9234 | 3.9261 |
227
+ | Warm CPU throughput / times real-time | 3.97x | 3.69x |
228
+ | Warm RTX 5090 throughput / times real-time | 55.17x | 53.45x |
229
+
230
+ **Protocol.** Deterministic replay balanced across all six preset voices, not a certified training-disjoint holdout or the official Open TTS leaderboard. ASR used Whisper large-v3 CTranslate2, FP16, beam 5, no VAD or reference prompt. Both sides used the same documented text normalization; endings and spelling variants were not forgiven. The MOS readers are automatic quality proxies, not human opinion scores.
231
+
232
+ **Timing.** Warm text preparation through returned waveform; 12 texts across six voices, three measured calls per text. CPU: Ryzen 7 7800X3D, two threads. GPU: RTX 5090. Loading, queueing, encoding and disk writes were excluded. These are not cold-start latencies or competitor comparisons. Higher times real-time is faster.
233
+
234
+ **Scope.** These figures do not evaluate Q1500, the combined profile, event tags or Tuner quality. The enhancer differences on the two quality readers did not establish a clear aggregate gain or loss. Both readers internally use 16 kHz, so they do not directly score bandwidth above 8 kHz.
235
+
236
+ Machine-readable provenance for the featured audio and reference numbers: [presentation evidence](assets/presentation-evidence.json).
237
+
238
+ </details>
239
+
240
+ ### Download checks, not just demo checks
241
+
242
+ The November 10 anonymous CPU check downloaded the public Mini repository without a token, verified all **445 inventoried files**, and rendered ordinary speech plus each of the four event tags. The returned audio matched the corresponding qualified outputs exactly. This establishes a functional check on that published revision, not a new perceptual-quality score or universal platform certification.
243
+
244
+ ## Small core. Explicit extras.
245
+
246
+ <p align="center"><img src="assets/mini-components.svg" alt="Original Mini: 34.31M core parameters. Add 3.80M for Cassidy delivery, 2.85M for enhancement, or both for 40.96M total. Event and Tuner components are separate." width="100%"></p>
247
+
248
+ | Speech configuration | Added learned parameters | Total learned speech parameters |
249
+ | :--- | ---: | ---: |
250
+ | Original Mini | 0 | **34,310,790** |
251
+ | Mini + Cassidy Q1500 | 3,801,977 | **38,112,767** |
252
+ | Mini + 48 kHz enhancer | 2,846,722 | **37,157,512** |
253
+ | Cassidy Q1500 + enhancer | 6,648,699 | **40,959,489** |
254
+
255
+ The base weight file is **138,388,547 bytes** (138.39 MB, decimal), at `core/baseline/models/mini/weights.pt`. Core parameters, weight-file size, complete download, installed dependencies and runtime memory are different quantities. Pronunciation assets and fixed voice-conditioning assets are not included in core parameter counts.
256
+
257
+ The event beta separately adds a **13,531,650-parameter Vocos decoder** and **25,600 learned control values**. Tuner enrollment uses separately acquired encoders; their footprint is not included in Mini's core count. See each component's guide for setup and terms.
258
+
259
+ ## A family, not one size for every job
260
+
261
+ | | Mini v3 | Micro v3 | Atom v3 |
262
+ | :--- | :--- | :--- | :--- |
263
+ | Core parameters | **34.31M** | **11.79M** | **157.19K** |
264
+ | Place in the family | Speech core with optional feature components | Smaller neural speech core | Ultra-compact synthesis core |
265
+ | Preset English voices | 6 | 6 | 6 |
266
+ | Native output | 24 kHz | 24 kHz | 24 kHz |
267
+ | Get the model | **You are here** | [Micro](https://huggingface.co/owensong/Inflect-Micro-v3) | [Atom](https://huggingface.co/owensong/Inflect-Atom-v3) |
268
+
269
+ Micro and Atom are separate models, not Mini with its optional components removed. Their capabilities and tradeoffs are documented on their own cards. Historical Preview repositories remain available and unchanged.
270
+
271
+ ## Before you build
272
+
273
+ Mini is a six-preset **English** model. Difficult pronunciation, numerals, long passages and delivery can still fail; review generated speech before using it. The 48 kHz option is enhancement, not a native-48 kHz backbone. Q1500 is Cassidy-only, and the four event performances are fixed. **Excited, angry, whisper and other research style tags are not supported in this release.** The experimental General frontend is not the default.
274
+
275
+ Use this release for **noncommercial evaluation** under the exact [Inflect Preview Evaluation License 1.0](LICENSE.md). Commercial use, redistribution and paid or free hosted services require separate written permission under those terms. These terms do not override upstream licenses or grant third-party voice or identity rights.
276
+
277
+ <details>
278
+ <summary><strong>Package integrity, licenses and acknowledgments</strong></summary>
279
+
280
+ `core/` contains the speech package; `extras/events/` and `extras/tuner/` contain independent optional components. The [outer release inventory](release-manifest.json), [Core inventory](core/assembly-manifest.json) and nested component manifests record separate compatibility identities. This presentation update does not change speech weights or runtime code.
281
+
282
+ Read the [top-level notices](THIRD_PARTY_NOTICES.md), [Core license scope](core/LICENSE_SCOPE.json), [Core notices](core/THIRD_PARTY_NOTICES.md), and the notices supplied with each optional component. Applicable [StyleTTS2 MIT](core/LICENSES/StyleTTS2-MIT.txt), [Apache-2.0](core/LICENSES/Apache-2.0.txt) and [NVIDIA REUSE OpenMDW-1.1](core/LICENSES/NVIDIA_REUSE_OpenMDW-1.1.txt) texts are retained. The REUSE teacher is not bundled. The event decoder has its own MIT notice. Tuner encoder terms include Apache-2.0 and an embedded CC BY-SA 3.0 speaker encoder; their obligations remain separate from Inflect's terms.
283
+
284
+ Private recordings and development recipes are not part of the public package. Required attribution is preserved without presenting a training recipe.
285
+
286
+ </details>
287
+
288
+ ---
289
+
290
+ <p align="center"><strong>Make something worth hearing.</strong><br>
291
+ <a href="#run-locally">Run Mini locally</a> &nbsp; / &nbsp;
292
+ <a href="https://github.com/owenawsong/Inflect">Explore Inflect</a> &nbsp; / &nbsp;
293
+ <a href="https://discord.gg/CVJYedvzvp">Join the conversation</a> &nbsp; / &nbsp;
294
+ <a href="https://huggingface.co/owensong/Inflect-Mini-v3/discussions">Feedback and issues</a></p>
release-manifest.json CHANGED
@@ -7,8 +7,8 @@
7
  "bytes": 4232
8
  },
9
  "README.md": {
10
- "sha256": "2973ac4dbce5811a9e27802b7240a349ecb8af11090850bb6eb5408511d52f43",
11
- "bytes": 18252
12
  },
13
  "THIRD_PARTY_NOTICES.md": {
14
  "sha256": "b22ade9c90934593770a39aa32c72f62c23f3ad17bc3a8473ab5ab9a1df1f660",
 
7
  "bytes": 4232
8
  },
9
  "README.md": {
10
+ "sha256": "24f908e945871ff299e3e13d0562fd8d1d63aea49c19cf7794c38d4e74291456",
11
+ "bytes": 19034
12
  },
13
  "THIRD_PARTY_NOTICES.md": {
14
  "sha256": "b22ade9c90934593770a39aa32c72f62c23f3ad17bc3a8473ab5ab9a1df1f660",