infosave/phonon-40m-multilingual-cmf

Phonon 40M Multilingual — CMF

This is a CMF distribution of Gradium's Phonon 40M Multilingual. It supports English, French, German, Spanish, and Portuguese and writes mono 24 kHz WAV audio.

The package deliberately does not claim a new quantization or a native reimplementation of Phonon. It keeps the upstream model.q8.gguf byte for byte, together with its config.json, tokenizer.json, and default-voice.safetensors, in integrity-checked CMF U8 entries. Cortiq verifies those entries, materializes them in a private temporary directory, and uses Gradium's maintained ptts runtime for the model-specific flow and Mimi-decoder inference path.

Published files

file	bytes	SHA-256
phonon-40m-multilingual-q8.cmf	75,380,093	c622ed46bf67a2f50965d4f5b7ee502b8d8580105f60530148b83b2cbf9a743b

The immutable artifact release is commit fcbc781f9108da6ea6c3a90b00e5c25875faac52.

Five short deterministic smoke renders (--temperature 0 --seed 42) are in samples/: English, French, German, Spanish, and Portuguese. They confirm that the packaged runtime wrote mono 24 kHz WAV output for every advertised normalization language; they are not a quality benchmark.

Use now

Use the public test Space for verified synthesis today. It runs the artifact with Gradium's maintained ptts runtime and exposes language, rate, temperature, seed, and CFG controls.

Cortiq CLI 0.8.36

cortiq-cli 0.8.36 contains the reviewed Phonon CMF bridge. Install the published crate (the default features include the phonon bridge), then verify the container before synthesis:

cargo install cortiq-cli --version 0.8.36 --locked

# After downloading the exact CMF artifact below:
cortiq verify phonon-40m-multilingual-q8.cmf
cortiq tts phonon-40m-multilingual-q8.cmf "A calm, clear sentence." \
  --lang en --out speech.wav

Prebuilt macOS, Linux, Windows, Android, and FFI archives are attached to the GitHub release v0.8.36. cortiq tts requires an explicit --lang of en, fr, de, es, or pt; it does not use automatic language detection.

To download the immutable CMF artifact for archival or integrity verification:

python -m pip install -U "huggingface_hub[cli]"
hf download infosave/phonon-40m-multilingual-cmf \
  phonon-40m-multilingual-q8.cmf --revision fcbc781f9108da6ea6c3a90b00e5c25875faac52 \
  --local-dir .

Useful controls

cortiq tts phonon-40m-multilingual-q8.cmf "A calm, clear sentence." \
  --lang en --temperature 0.3 --seed 4242424242424242 \
  --duration-delta -0.25 --out fast.wav

• --temperature controls flow-noise variation; 0.3 is the upstream default and 0 is deterministic for a fixed seed.
• --seed makes the stochastic path reproducible.
• --duration-delta is the model's native rate conditioner: negative values ask for faster speech and positive values ask for slower speech. It is not a pitch-changing post-processing resampler.
• --guidance, --eos-threshold, and --max-tokens-per-chunk are available for advanced use.
• --voice-file accepts only an already computed Phonon-compatible .safetensors voice latent. Recording-based voice cloning is not exposed: the needed speaker codec is not distributed by the upstream model.

Only the upstream default voice is in this artifact. Do not assume that a voice from Gradium's separate voice catalog may be redistributed here; check that voice's own terms first.

Contents and integrity

CMF entry	upstream file	purpose
phonon.model_q8_gguf	model.q8.gguf	upstream Q8 model payload
phonon.config_json	config.json	model configuration
phonon.tokenizer_json	tokenizer.json	normalization/tokenization data
phonon.default_voice_safetensors	default-voice.safetensors	default precomputed voice latent

cortiq verify validates the CMF envelope, directory, and the hash of every stored entry before the TTS backend is invoked. The companion Space also validates this restricted CMF contract before extracting its assets.

Provenance and license

The source checkpoint is gradium/phonon-40m-multilingual at revision 1537ce46c64f42903e0d608b132ccb073dac46d2. It is a gated upstream model: you must accept its terms to obtain the source checkpoint.

The source payload SHA-256 values preserved in the CMF provenance are:

e312ab694d5ec1f79ccea520a2eeec989969ce07e9b4189ae7e9ad8e035ef82e  config.json
96fa067813244e0c6c5997b98ec2596b65ec8574b14916b48e97db056ff97e27  model.q8.gguf
a5e8c6060bf927ed367314f21355148fdf4acff6fadb94e49c86bc7182449433  tokenizer.json
dc537478eaadc1a5257fcda2e3501ab81e14c111a5d95a5838f350db2afd0f1a  default-voice.safetensors

The weights and this derived redistribution are under CC BY-NC 4.0: attribution is required and commercial use is not permitted. The release includes the upstream license/notice material and preserves this provenance. The Cortiq bridge uses the upstream ptts code under its MIT-or-Apache-2.0 terms; that does not change the model-weight license.

No quality, latency, or voice-similarity claim is made here without a reproducible measurement accompanying the published artifact.