edgardozavala/phonon-tts-31m

Lightweight, CPU-powered English Text-to-Speech model based on Gradium Phonon 31M.

Public
2 runs

Run time and cost

This model runs on CPU hardware. We don't yet have enough runs of this model to provide performance information.

Readme

Phonon TTS 31M

Lightweight, high-quality English text-to-speech powered by Gradium Phonon 31M, running on CPU only. No GPU, no PyTorch, no CUDA.

Features

  • 31M parameter model, ~72 MB with all 38 voices baked into the image
  • CPU-only inference, ~8x faster than realtime
  • 24 kHz mono output, native WAV (16-bit PCM)
  • Optional MP3 at 64, 96, 128 or 192 kbps
  • 38 English voices: 25 standard (US and UK) and 12 character voices
  • No network access at prediction time
  • Apache-2.0 weights and voices

Model

  • Checkpoint: gradium/phonon-31m-en at commit 8f7a37c9… (Apache-2.0)
  • Voices: gradium/phonon-voices at commit 0ee17205… (Apache-2.0)
  • Runtime: gradium-ai/xn-ptts (MIT OR Apache-2.0)
  • Language: English

Inputs

Parameter Type Default Description
text string required English text to synthesize, 1–4096 characters
voice string default One of the 38 voices listed below
format string wav wav or mp3
bitrate string 128k 64k, 96k, 128k, 192k. Only applies to mp3; ignored for WAV
language string en English only; kept for API compatibility

Voice names are case-insensitive. default is the checkpoint’s built-in voice.

Voices

American English (US)

Female Male
Delia Easton
Everly Micah
Hadley Miles
Harper Preston
Juniper Russell
Zoey Trevor

British English (UK)

Female Male
Bryony Alfie
Elodie-Rose Edwin
Freya Hugh
Imogen Spencer
Lottie Teddy
Maeve Toby
Pippa

Character voices

Azgorath · Captain Silas Blackwater · Fairy Godmother · Fennick the Young Elf · Grub the Ogre · Master Aldric · Morwenna of the Deep · Old Mortlake · Sir Bramwell · The Snow Queen · The Thing Under the Stairs · Vyrmathrax the Ancient

Output

WAV (default) is the model’s native output, returned untouched: RIFF/WAVE, PCM signed 16-bit little-endian, 24 kHz, mono. No recompression, no resampling.

MP3 is encoded from that WAV with FFmpeg (libmp3lame) at 48 kHz and the requested bitrate. MP3 cannot exceed 160 kbps at 24 kHz, so 48 kHz is used to make 192 kbps real instead of silently capped. The intermediate WAV is deleted after conversion, and every MP3 is checked with ffprobe before it is returned.

Limitations

  • English only.
  • Maximum 4096 characters per request; longer text must be split by the caller.
  • One prediction at a time per instance (concurrency.max: 1).
  • No streaming output; the complete file is returned.
  • No voice cloning or custom voice uploads.
  • MP3 at 48 kHz; WAV is 24 kHz.

License

  • Model weights: Phonon 31M, Apache-2.0, by Gradium.
  • Voices: gradium/phonon-voices, Apache-2.0.
  • Runtime: gradium-ai/xn-ptts, MIT OR Apache-2.0.

Phonon 40M and the multilingual variants are CC BY-NC 4.0 (non-commercial) and are not part of this model.

Model created
Model updated