Preview: user-facing v0.1 model card

README.md

@@ -18,259 +18,204 @@ language:
   - nl
   - sv
   - da
-base_model: ResembleAI/chatterbox
-base_model_relation: finetune
 ---
 
 # loudr-1
 
-**loudr-1** is a text-to-speech model that runs on your own machine. Twenty
-voices across ten languages, a new voice cloned from ten seconds of audio, and
-the same speech tokens from five implementations of the engine. No account, no
-network, no telemetry.
+**Natural-sounding text-to-speech with twenty voices, ten languages and voice
+cloning.**
 
-It is the checkpoint for [`loudkit`](https://github.com/loudreader/loudkit):
-one file, mixed precision by module, with the algorithm's own settings recorded
-inside it so no backend has to guess them.
+loudr-1 runs on your own hardware through
+[loudkit](https://github.com/pepinu/loudkit). Download it once and work offline
+from Python, Swift, Go, Rust or TypeScript with PyTorch, ONNX Runtime or CoreML.
 
-## Listen first
+[**Hear all 20 voices**](https://pepinu.github.io/loudkit/demo/) |
+[**Open in Colab**](https://colab.research.google.com/github/pepinu/loudkit/blob/main/notebooks/loudkit_quickstart.ipynb) |
+[**GitHub**](https://github.com/pepinu/loudkit) |
+[**Documentation**](https://pepinu.github.io/loudkit/)
 
-Every voice has a playable sample on the
-[demo page](https://loudreader.github.io/loudkit/demo/). `joe` and `joe`
-read the identical sentence from *Alice in Wonderland*, so they are a real A/B
-of the model rather than of the material. Sample provenance, down to the seed,
-is recorded per voice in
-[provenance.json](https://github.com/loudreader/loudkit/blob/main/docs/voices/roster/provenance.json).
+## Listen
 
-## Links
+**Joe**
 
-- [GitHub](https://github.com/loudreader/loudkit): the engine, five
-  implementations
-- [Documentation](https://loudreader.github.io/loudkit/): guides and reference
-- [Demo](https://loudreader.github.io/loudkit/demo/): hear the voices
-- [Voices](https://github.com/loudreader/loudkit/blob/main/VOICES.md): the full
-  roster, with source and licence per voice
-- [Measured performance](https://loudreader.github.io/loudkit/benchmarks/):
-  speed per device and per execution provider, with the command that reproduces
-  each number
+<audio controls src="https://huggingface.co/loudreader/loudr-1/resolve/main/samples/joe.opus"></audio>
 
-## Usage
+**Kathleen**
+
+<audio controls src="https://huggingface.co/loudreader/loudr-1/resolve/main/samples/kathleen.opus"></audio>
+
+Both voices read the same passage from *Alice's Adventures in Wonderland*.
+[Open the gallery](https://pepinu.github.io/loudkit/demo/) to compare every
+shipped voice with the enrollment reference used to create its profile.
+
+> English is the only language we could evaluate ourselves by ear. We do not
+> speak the other nine languages well enough to judge their naturalness
+> reliably. If you do, please listen and share what sounds good or wrong.
+> Feedback from native speakers is very welcome.
+
+## Start in Python
 
 ```bash
 pip install "loudkit[torch,audio,hub]"
 ```
 
-The package is published alongside each release. If pip cannot find it yet,
-follow the [install instructions on
-GitHub](https://github.com/loudreader/loudkit#hear-it).
-
 ```python
 import loudkit as lk
 
 engine = lk.load("loudreader/loudr-1")
-narrator = lk.voice("joe", repo="loudreader/loudr-1")
+voice = lk.voice("joe", repo="loudreader/loudr-1")
 
-engine.synthesize("Hello from loudkit.", narrator, seed=7).save("hello.wav")
+engine.synthesize("Hello from loudkit.", voice, seed=7).save("hello.wav")
 ```
 
-Or from the shell:
+The first run downloads the 747 MB synthesis checkpoint and the voices. Later
+runs use the local cache. The same path from a shell is:
 
 ```bash
-loudkit speak --checkpoint loudreader/loudr-1 --voice joe "Hello from loudkit."
+loudkit speak --checkpoint loudreader/loudr-1 --voice joe \
+  "Hello from loudkit." -o hello.wav
 ```
 
-The first synthesis downloads this checkpoint (1.27 GB); everything after that
-runs from the cache. Cloning a voice takes ten seconds of clean audio and
-produces a ~300 KB file of tensors:
+To enroll a voice that you own or have permission to use:
 
 ```python
 mine = lk.enroll("my-recording.wav", "loudreader/loudr-1", name="my-voice")
 mine.save("voices/my-voice.safetensors")
 ```
 
-## What it is
-
-Two stages. A **token generator**, a Llama-architecture decoder (16 layers,
-hidden 1024, GQA with 4 KV heads, 145M parameters), writes discrete speech
-tokens at 25 Hz, autoregressively. A **renderer**, a conditional flow-matching
-mel decoder plus a HiFiGAN vocoder, turns those tokens into a 24 kHz waveform
-in one parallel pass.
-
-The two have opposite shapes, so every backend treats them differently. On
-Apple silicon the generator is faster on the CPU and the renderer is faster on
-the GPU.
-
-## How it was made
-
-Distilled from [Chatterbox](https://github.com/resemble-ai/chatterbox) (MIT,
-Resemble AI) in three steps, each measured rather than assumed:
-
-1. **Guidance distillation.** Classifier-free guidance runs the flow estimator
-   twice per step. The student learns to emit the combined velocity directly,
-   in one pass.
-2. **Step distillation.** Six Euler integration steps to two, by training the
-   student's two-step rollout endpoint against the teacher's six-step endpoint.
-3. **Quantisation-aware training** for the estimator, after post-training
-   quantisation failed: int8 PTQ dropped velocity-field correlation from 0.980
-   to 0.924, and QAT with straight-through estimators recovered it.
-
-Distillation targeted seven product languages. The release ran against a frozen
-evaluation set of 11 languages × 50 sentences = 550 items, cut
-deterministically from the corpus (hash-sorted, stratified by length) with
-voices rotated so a gate cannot be passed by one speaker. Four gates, applied
-per language rather than in aggregate, because an average hides the one
-language that broke: teacher/student KL, length ratio, early end-of-speech
-rate, and word error rate against an ASR model. The shipped student cleared all
-four in every product language. Thresholds and per-language values are internal
-training records; this card states only the verdict.
-
-**Those four gates are metrics, not ears.** Clearing them means the student
-tracks the teacher and does not truncate or ramble. It does not mean a language
-sounds good. Listening quality has been evaluated for **English only**.
-
-### The five language counts, and why they differ
-
-This project carries five counts. Conflating any two of them produces a false
-claim:
-
-| count | what it describes |
-| --- | --- |
-| **23** | the teacher model's language coverage |
-| **7** | the languages distillation optimised for |
-| **11** | the languages the release gates were measured on (× 50 sentences) |
-| **12** | the languages the text frontend verbalises numbers, dates and times in |
-| **10** | the languages voices actually ship for |
-
-Separately, the tokenizer inherits **31** two-letter language tags. That is a
-vocabulary fact, not a capability: the frontend refuses anything outside the 12
-it has grammars for.
-
-## Precision, per module
-
-Mixed on purpose, and each decision has a measurement behind it:
-
-| module | precision | why |
-|---|---|---|
-| token generator | fp16 | teacher-forced median KL 1.29e-06 against fp32, top-1 agreement 99.9%. A sampling decision boundary annihilates error below it. |
-| flow estimator | fp16 | mel correlation 0.999999, waveform 0.9987 |
-| flow encoder | **fp32** | fp16 here gives mel correlation 0.619 and +22 dB of high-frequency energy |
-| vocoder | **fp32** | fp16 produces an audible tone at Nyquist: the source module accumulates phase to ~1400 cycles, where fp16 resolution is coarser than the per-sample increment |
-
-The same numerical change that is free for a discrete-token decoder is
-destructive for a continuous-output regressor, because there is no decision
-boundary downstream to absorb it.
-
-## Determinism
-
-Same text, voice and seed give a **bit-identical waveform** on a given build,
-device and backend. Every backend implements one counter-based sampling law, so
-the same seed makes the same token decisions within the envelope documented in
-[the identity
-contract](https://github.com/loudreader/loudkit/blob/main/docs/reference/IDENTITY-CONTRACT.md).
-That is not free: `torch.multinomial` returns different samples for an
-identical probability vector on x86 and arm64.
-
-**Not** bit-identical across backends or across releases. Different hardware
-sums floating-point numbers in different orders, and engine changes re-base the
-goldens under a versioned contract. Token identity across backends holds *at
-matched precision*: fp16 in the generator flips roughly one token in a
-thousand, and one flip re-routes everything after it.
+The reusable profile is about 150 KB. Install
+`loudkit[torch,audio,enroll,hub]` for enrollment.
+
+## Choose your runtime
+
+The repository contains all supported formats, but the downloader fetches only
+the runtime you select. Add `--with-cloning` when the installation also needs
+enrollment.
+
+| path | command | download |
+|---|---|---:|
+| Python, synthesis | `loudkit download loudreader/loudr-1 --for torch` | 750 MB |
+| Python, with cloning | `--for torch --with-cloning` | 1.28 GB |
+| JS, Go or Rust with ONNX | `--for onnx` | 2.60 GB |
+| ONNX, with cloning | `--for onnx --with-cloning` | 3.13 GB |
+| Swift or Python with CoreML | `--for coreml` | 1.16 GB |
+| CoreML, with cloning | `--for coreml --with-cloning` | 1.69 GB |
+
+Add `--local-dir loudr-1` to create a portable directory instead of using the
+shared cache. The synthesis checkpoint will be at
+`loudr-1/loudr-1.safetensors`.
+
+### Measured speed
+
+| path | hardware | real-time factor |
+|---|---|---:|
+| PyTorch with CUDA graphs | RTX 3090 | 7.47x |
+| PyTorch with CUDA graphs | Jetson Orin Nano | 1.83x |
+| split PyTorch engine | Apple M3 Pro | 3.43x |
+| ONNX Runtime, CPU provider | Apple M3 Pro | 1.21x |
+| PyTorch CPU reference | Apple M3 Pro | 0.33x |
+
+Higher is faster, and 1.0x means real time. ONNX Runtime on the measured M3 Pro
+CPU is faster than real time. The PyTorch CPU reference path on the same machine
+is not.
+
+For batched workloads, the token generator reaches 20.1x aggregate throughput
+at batch 1 and 153.1x at batch 64 on the RTX 3090. The highest measured result
+is 170.8x on an A100 at batch 64. These are generator-only throughput numbers,
+not single-request latency or end-to-end RTF. See the
+[benchmark report](https://pepinu.github.io/loudkit/benchmarks/) for commands,
+hardware and caveats.
+
+## What ships
+
+| artefact | size | used by |
+|---|---:|---|
+| `loudr-1.safetensors` | 747 MB | synthesis |
+| `loudr-1-enrollment.safetensors` | 523 MB | PyTorch enrollment |
+| `ve.safetensors` | 5.7 MB | PyTorch enrollment |
+| `onnx/` | 2.38 GB | nine graphs: six synthesis, three enrollment |
+| `coreml/` | 941 MB | six packages: three synthesis, three enrollment |
+| `voices/` | 3.1 MB | twenty voice profiles |
+| `samples/` | 108 KB | the two players above |
+| `tokenizer.json` | 70 KB | text processing |
+
+Synthesis and enrollment are separate so users who only need speech generation
+do not download the enrollment weights. ONNX and CoreML use their own enrollment
+graphs. loudkit also verifies that paired model files came from the same source
+checkpoint.
+
+## Voices and consent
+
+The release includes two profiles for each of these languages: English,
+Spanish, French, German, Italian, Polish, Portuguese, Dutch, Swedish and Danish.
+
+The profiles were built from recordings donated for speech technology or from
+CC0 and CC-BY speech corpora. No scraped celebrity voices ship with the model.
+[The full roster](https://github.com/pepinu/loudkit/blob/main/VOICES.md) records
+the source, licence and consent basis for every profile. The
+[voice gallery](https://pepinu.github.io/loudkit/demo/) provides a generated
+sample and enrollment preview for all twenty.
+
+The source enrollment WAVs are not redistributed in the model repository. Their
+digests, construction notes and the digests of every shipped profile and sample
+are recorded in
+[provenance.json](https://github.com/pepinu/loudkit/blob/main/docs/voices/roster/provenance.json).
+
+## Model lineage
+
+loudr-1 is derived from
+[Chatterbox](https://github.com/resemble-ai/chatterbox), released by Resemble AI
+under the MIT licence. We optimized it for faster local inference by profiling
+the full synthesis path, changing the signal flow, separating synthesis from
+enrollment, and adjusting graph boundaries and device placement for PyTorch,
+ONNX Runtime and CoreML.
+
+Release gates compare the implementations, check output length and early end of
+speech, and run ASR-based checks per measured language. These checks catch
+mechanical regressions. They do not replace listening by native speakers.
+
+## Reproducibility
+
+For a fixed build, device and backend, the same text, voice and seed produce the
+same waveform. Across devices or backends, loudkit checks the token stream and
+keeps waveform differences inside measured correlation bands. Floating-point
+execution means that waveforms are not promised to be byte-identical across
+different runtimes.
+
+The exact contract and current measurements are in the
+[identity contract](https://github.com/pepinu/loudkit/blob/main/docs/reference/IDENTITY-CONTRACT.md)
+and [measured parity report](https://pepinu.github.io/loudkit/parity-measured/).
+
+## Before you ship
+
+- Long passages are rendered in windows of about ten seconds. Sentence joins
+  can occasionally be audible.
+- Difficult punctuation, numbers and abbreviations can change pronunciation or
+  prosody.
+- Voice cloning requires consent. A recording being public does not grant
+  permission to clone the speaker.
+- Saved WAVs and server responses include an unsigned C2PA Content Credentials
+  manifest by default. It records the model, voice, seed, backend and audio
+  digest in a machine-readable form.
+
+Read [Responsible use](https://huggingface.co/loudreader/loudr-1/blob/main/RESPONSIBLE_USE.md)
+before exposing enrollment to other people.
 
 ## Intended use
 
-Narration, accessibility, localisation, research: any application that speaks
-text. Local by design, with no account, no network and no telemetry.
-
-## Limitations
-
-- **Model quality validated for English only.** The architecture is
-  multilingual and the training data was not restricted to English. The text
-  frontend can *spell out* other languages graphemically, including Polish,
-  which has a full numbers→words/anglicism/respelling pipeline. Only English
-  has been listening-evaluated. Token parity across the five implementations is
-  measured for every language; how good the result sounds is not. A language
-  nobody has listened to will be read, not spoken well. Listen before you ship
-  it.
-- **~10 seconds per window.** Longer passages are split at sentence boundaries.
-  The library does this, and where the breaks fall is audible.
-- **Marking is metadata only; there is no audio watermark.** From 2026-08-02
-  the EU AI Act (Article 50) requires synthetic audio to carry a
-  machine-readable marking. loudkit ships the **C2PA claim-only manifest**
-  written into every saved WAV and every server response
-  ([provenance](https://github.com/loudreader/loudkit/blob/main/docs/reference/provenance.md)).
-  It is *unsigned*, so it is a disclosure aid, not proof of origin. An in-audio
-  watermark was evaluated and not shipped: its detection statistic was not
-  separable from the null. Any re-encode strips the metadata and a re-synthesis
-  defeats it entirely, so disclosure remains the user's responsibility. See
-  [RESPONSIBLE_USE.md](https://huggingface.co/loudreader/loudr-1/blob/main/RESPONSIBLE_USE.md).
-- **The generator dominates on Apple hardware.** The renderer is exported to
-  CoreML and fast. The generator is not, and runs natively.
-- Prosody on unusual input (heavy punctuation, numbers, abbreviations) is where
-  this model is weakest, as is common for the family.
-
-## Voices
-
-Twenty shipped voice profiles across ten languages: English, Spanish, French,
-German, Italian, Polish, Portuguese, Dutch, Swedish, Danish. Two per language,
-one of each gender. Each is an enrollment of a recording made or released for
-speech-technology use, enrolled by this project's own pipeline. None are
-Resemble AI artefacts, and none are cloned from scraped audio.
-
-The release policy is that **every voice carries its licence and its consent
-basis**. Ten of the twenty are personal donations recorded expressly for
-building TTS voices (CC0, via the OHF-Voice / NabuCasa voice-datasets
-collection). The rest come from CC0 speech databases (NST Swedish and Danish)
-or CC-BY-4.0 corpora (Kyutai tts-voices, CML-TTS, Multilingual LibriSpeech,
-which are LibriVox-derived recordings whose terms credit the readers).
-
-Donor names, sources, licences, consent basis, reference construction and
-sha256 for every profile live in
-[provenance.json](https://github.com/loudreader/loudkit/blob/main/docs/voices/roster/provenance.json),
-alongside speaker similarity wherever it was recorded (0.89–0.96 across the
-profiles that carry the measurement).
-
-**No cloned voice ships with this project.** Enrollment code is included so you
-can create one from a voice you have the right to use.
-
-## Risks, stated plainly
-
-This clones a voice from a few seconds of audio and reads arbitrary text in it,
-locally, with no gatekeeping. Those properties are the point, and they also
-make misuse cheap. Read
-[RESPONSIBLE_USE.md](https://huggingface.co/loudreader/loudr-1/blob/main/RESPONSIBLE_USE.md).
-Consent is not implied by a recording being public.
-
-## Training data and compute
-
-The distillation used speech generated by the teacher model, plus recordings
-made or released for speech-technology use for the reference voices (see
-[Voices](#voices)). The teacher's own training data is Resemble AI's and is not
-documented here. Anyone relying on provenance should read that project's terms
-rather than ours.
-
-Hardware: two consumer RTX 3090 GPUs. The full campaign (corpus generation,
-scaling probes, final training, evaluation) ran over roughly two weeks at about
-435–440 GPU-hours. Around half of that budget was corpus generation, which is
-teacher inference: producing the distillation data cost more than training the
-shipped student (~15 GPU-hours). The recipe involved more than distillation
-alone, and this card does not try to compress it. **A technical report covering
-everything that was done is pending.**
-
-## Citing
-
-If you use the model or the toolkit:
-
-```bibtex
-@software{loudkit2026,
-  title  = {loudkit: a portable text-to-speech engine},
-  author = {LoudReader},
-  year   = {2026},
-  url    = {https://github.com/loudreader/loudkit}
-}
-```
+loudr-1 is intended for local narration, accessibility, localisation, games,
+prototyping and speech research. It is not a voice-authentication system and
+must not be used for deceptive impersonation.
+
+## Training data
+
+The original Chatterbox training data is controlled by Resemble AI and is not
+documented by this project. The shipped voice profiles use recordings made or
+released for speech-technology use; their sources and licences are listed in
+the public roster.
 
 ## Licence
 
 [Apache-2.0](https://huggingface.co/loudreader/loudr-1/blob/main/LICENSE).
-Component licences:
+Upstream attributions and component licences are listed in
 [NOTICE](https://huggingface.co/loudreader/loudr-1/blob/main/NOTICE).

RESPONSIBLE_USE.md

@@ -60,13 +60,18 @@ fingerprint, the seed, and the SHA-256 binding it to the audio bytes. See
 because signing needs a certificate and that is the deployer's choice. It says
 what made the file, not who vouches for it.
 
-**That is the whole of the machine-readable marking. There is no audio
-watermark.**
+This is the machine-readable marking loudkit provides by default. Deployers can
+add their own signing policy and disclosure around it.
 
 ## What we ship, and what we will not
 
-We ship no voice cloned by us, from anyone, including the voices in our own
-product. The enrollment code is included so you can make your own, from a voice
+We ship twenty voice profiles enrolled from recordings donated for speech
+technology or released under terms that permit this use. Every profile has a
+named source, licence and consent basis in the public roster. We do not ship
+profiles made from private recordings or recordings published for an unrelated
+purpose without the speaker's permission.
+
+The enrollment code is included so you can make your own profile from a voice
 you have the right to use.
 
 Issues and pull requests asking for help with undisclosed impersonation, voice

SHA256SUMS

@@ -1,5 +1,6 @@
-26075de97f0dd09e56abf49cfede42828aafc90d025eae303102bb643a5d3d1c  loudr-1.safetensors
-0688ea295583be6497ec2883a7e55e3fa3463ae4cc3e4aa33fccce3249bf8e25  manifest.json
+688e87cddfc95a37b6705bda05ad95a7ddb219a4c0c232fa4b93ccadc42efe2c  loudr-1.safetensors
+dc825dcdec843cc0fb40ad3a0f051e869db97689a33712d1d11c009ce16af022  loudr-1-enrollment.safetensors
+795d019797197a0f01c722af516ac6521cda59c6a27d8d07828489175659a99f  manifest.json
 69632f47220a788a52ce2661d096453c5655e9bf25289d89a8d832c46ee07dbf  tokenizer.json
 f0921cab452fa278bc25cd23ffd59d36f816d7dc5181dd1bef9751a7fb61f63c  ve.safetensors
 78988a6769f90c209240f9f285eb9cdd760d1ba471cfa7d6e67db632599cfa1b  voices/darkman.safetensors
@@ -49,8 +50,10 @@ e7f2f756511fe047d0cd1997cdf2d42ba5c0c9fc8273c4450d25587e9e7aa77e  coreml/camp.ml
 759103649d9bf8d388b523b915e1eb9b285fb81e58fdef6d420d4c7e21053765  coreml/voice_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
 95b1a8b8c6b558fafee4a1c784076c8e099070c45958526c9a187c15de55d2d3  coreml/voice_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
 fbd810d98add4c3349f01891e1b7c6dc975070bfa83e78f8f8a115f5dcea3c55  coreml/voice_encoder.mlpackage/Manifest.json
-c399332def766b45b9c121515a72121aae648c214398df6effd11a376fb12fde  README.md
+dac906a2310982bb236b258db60f81141224087d54a3a66259332a74cb786c4a  README.md
 5d5671b95e9797856175dd08d5d2f20d21d72e8785aba077ee38eb378d432528  LICENSE
 fda3657fa4d457f9dc379c1ca8550d665eb972ac1c7a260fbe94821ef418ff28  NOTICE
-c1fe707141d6fda3f4daf53ca259eaef79067564847738843e91aa3bf9b90d18  RESPONSIBLE_USE.md
-6e0959d98112be647fc2f9c992de1e9977cf06ba3cfa623a6484095dfa5fe5d4  release.json
+87c35454a03b026a5a13093566ace245c339dc6af68c7f794f6ed322a731f753  RESPONSIBLE_USE.md
+83e0bf0a1f6b03e047de66871536c3ff77291041d0a3f2907b6990e580519fea  samples/joe.opus
+48615517c1d1d2d55dbb7f74f37982918236e33f079947459d56c116cd70b7a3  samples/kathleen.opus
+4058b467cd4a585bc1f39f1d236793181fef0b2e31dacdfb4cc8e31fb22f0aff  release.json

loudr-1-enrollment.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc825dcdec843cc0fb40ad3a0f051e869db97689a33712d1d11c009ce16af022
+size 522832312

loudr-1.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:26075de97f0dd09e56abf49cfede42828aafc90d025eae303102bb643a5d3d1c
-size 1269460068
+oid sha256:688e87cddfc95a37b6705bda05ad95a7ddb219a4c0c232fa4b93ccadc42efe2c
+size 746630956

manifest.json

@@ -1,18 +1,27 @@
 {
+ "artifact_role": "synthesis",
+ "chunking": {
+  "enabled": true,
+  "max_tokens": 255,
+  "prefix_tokens": 6,
+  "split_on": [
+   ". ",
+   "! ",
+   "? ",
+   "; ",
+   ", "
+  ]
+ },
  "dtype_map": {
   "s3gen.flow": "float32",
   "s3gen.flow.decoder.estimator": "float16",
   "s3gen.mel2wav": "float32",
-  "s3gen.speaker_encoder": "float32",
-  "s3gen.tokenizer": "float32",
   "t3": "float16"
  },
  "dtype_rationale": {
   "s3gen.flow": "EXP-011: whole-flow fp16 destroys output (mel corr 0.619, +22 dB HF); damage is in the encoder",
   "s3gen.flow.decoder.estimator": "EXP-011: fp16 mel corr 0.999999, wave corr 0.9987, no new HF energy",
   "s3gen.mel2wav": "EXP-003: fp16 produces audible ~12 kHz tone (NSF cumulative phase accumulator). Non-negotiable fp32.",
-  "s3gen.speaker_encoder": "enrollment support; never measured under reduced precision -> fp32",
-  "s3gen.tokenizer": "enrollment support; never measured under reduced precision -> fp32",
   "t3": "EXP-010: fp16 teacher-forced median KL 1.29e-06 vs fp32, top-1 99.9%; bf16 rejected (no ANE datapath)"
  },
  "eos_floor": {
@@ -56,6 +65,27 @@
  },
  "n_cfm_timesteps": 2,
  "name": "loudkit-v0.1",
+ "postprocess": {
+  "ceiling_slack_tokens": 40,
+  "ceiling_speech_per_text_token": 4.0,
+  "desperation_min_text_tokens": 10,
+  "desperation_speech_per_text_token": 4.5,
+  "echo_strong_eos_probability": 0.1,
+  "echo_strong_max_tail": 30,
+  "echo_strong_min_position_pct": 68,
+  "echo_weak_eos_probability": 0.003,
+  "echo_weak_max_tail": 16,
+  "echo_weak_min_position_pct": 85,
+  "ended_tail_blip_max": 2,
+  "ended_tail_keep": 5,
+  "ended_tail_silence_run": 6,
+  "ended_tail_word_max": 10,
+  "filler_max_speech_after_run": 10,
+  "filler_min_eos_probability": 0.05,
+  "mode": "trim",
+  "trailing_filler_threshold": 0.7,
+  "trailing_silence_run_tokens": 12
+ },
  "recipe_version": "loudkit-1",
  "runtime_requirements": {
   "allow_tf32_cudnn": false,
@@ -151,6 +181,15 @@
   "stop": 6562
  },
  "speech_vocab_size": 8194,
+ "split": {
+  "roles": {
+   "enrollment": "loudr-1-enrollment.safetensors",
+   "synthesis": "loudr-1.safetensors"
+  },
+  "source_payload_sha256": "09957204231efbdf99e73220fdb6abfa2cf313f124184e5bcf36bfc4feba17eb",
+  "source_tensor_count": 2574,
+  "source_tensor_names_sha256": "d0f4888a6b123444a3ff5084bb1b6da8574af04c25af80cd06342ade33c05118"
+ },
  "t3_config": "multilingual",
  "t3_keep_layers": [
   0,
@@ -171,7 +210,9 @@
   29
  ],
  "t3_llama_config_registered_as": "Llama_520M",
- "tensor_payload_sha256": "09957204231efbdf99e73220fdb6abfa2cf313f124184e5bcf36bfc4feba17eb",
+ "tensor_count": 1533,
+ "tensor_payload_sha256": "9972afd402cd7349fb5320f30855fab1d2d01b201cfb46ae20868193bf211ec5",
+ "tokenizer_sha256": "69632f47220a788a52ce2661d096453c5655e9bf25289d89a8d832c46ee07dbf",
  "torch_version_at_pack": "2.6.0",
  "weight_norm": {
   "folded": true,

release.json

@@ -2,13 +2,18 @@
  "profile": "full-0.1",
  "checkpoint": {
   "path": "loudr-1.safetensors",
-  "sha256": "26075de97f0dd09e56abf49cfede42828aafc90d025eae303102bb643a5d3d1c",
-  "bytes": 1269460068
+  "sha256": "688e87cddfc95a37b6705bda05ad95a7ddb219a4c0c232fa4b93ccadc42efe2c",
+  "bytes": 746630956
+ },
+ "enrollment_checkpoint": {
+  "path": "loudr-1-enrollment.safetensors",
+  "sha256": "dc825dcdec843cc0fb40ad3a0f051e869db97689a33712d1d11c009ce16af022",
+  "bytes": 522832312
  },
  "manifest": {
   "path": "manifest.json",
-  "sha256": "0688ea295583be6497ec2883a7e55e3fa3463ae4cc3e4aa33fccce3249bf8e25",
-  "bytes": 4599
+  "sha256": "795d019797197a0f01c722af516ac6521cda59c6a27d8d07828489175659a99f",
+  "bytes": 5635
  },
  "tokenizer": {
   "path": "tokenizer.json",
@@ -263,8 +268,8 @@
  ],
  "readme": {
   "path": "README.md",
-  "sha256": "c399332def766b45b9c121515a72121aae648c214398df6effd11a376fb12fde",
-  "bytes": 11908
+  "sha256": "dac906a2310982bb236b258db60f81141224087d54a3a66259332a74cb786c4a",
+  "bytes": 8274
  },
  "license": {
   "path": "LICENSE",
@@ -278,8 +283,20 @@
  },
  "responsible_use": {
   "path": "RESPONSIBLE_USE.md",
-  "sha256": "c1fe707141d6fda3f4daf53ca259eaef79067564847738843e91aa3bf9b90d18",
-  "bytes": 3699
+  "sha256": "87c35454a03b026a5a13093566ace245c339dc6af68c7f794f6ed322a731f753",
+  "bytes": 4022
  },
+ "samples": [
+  {
+   "path": "samples/joe.opus",
+   "sha256": "83e0bf0a1f6b03e047de66871536c3ff77291041d0a3f2907b6990e580519fea",
+   "bytes": 52187
+  },
+  {
+   "path": "samples/kathleen.opus",
+   "sha256": "48615517c1d1d2d55dbb7f74f37982918236e33f079947459d56c116cd70b7a3",
+   "bytes": 55798
+  }
+ ],
  "verified": true
 }