TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #102

Generate speech in almost any language with control over age, accent, pitch and style

Open text-to-speech models usually cover a handful of languages and one flavor of accent. This small independent release goes the other way: hundreds of languages, thirteen Arabic dialects, and a voice you describe in words.

MODELice-012-audio
PUBLISHEDSeptember 3, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYAUDIO

Field notes

01What it does

ICE-012 Audio converts text into speech and lets you shape the voice without recording anything. You choose the language by name or code, set male or female, and add plain-word modifiers like "elderly, low pitch" or "teenager, british accent". Speed runs from half to double pace, and output streams as it is generated, so long passages start playing before they finish rendering. For a specific voice, you pass a reference WAV and optionally its transcript, and the model clones that speaker for new text.

The language list is unusually wide: 590 names, including regional Arabic dialects such as Egyptian, Saudi, Moroccan, and Levantine variants, which mainstream TTS models usually flatten into one generic Arabic. The author credits a new acoustic adapter that conditions the codec embeddings before the main network, which is the technical reason a model this size can cover so much ground. The weights are Apache 2.0.

02How to try it

No demo Space yet, but the model is small enough to run on a single consumer GPU: about 700 million parameters, one safetensors file, loadable through transformers with remote code enabled. The model card has a quick-start snippet and the repo includes an inference script and a notebook. A good first test is the same English sentence rendered three ways, say a young adult with a British accent, an elderly low-pitched narrator, and a whisper, then one line in a language your current TTS setup cannot handle.

03Caveat

Coverage this wide is a claim to test, not a guarantee: expect quality to vary across the long tail of languages, and the card only ships a few Arabic and English samples. This is an independent developer's release rather than a lab's, and the command-line scripts named in the README are not in the repo, so start from the quick-start code and the included inference script. As always, only clone voices you have permission to use.

04What you can do with it

  • Narrate the same script in several regional Arabic dialects for local audiences.
  • Prototype a character voice by describing it instead of casting or recording it.
  • Clone your own voice from a clean clip and read new text with it.
  • Add streaming speech to an app in a language the big TTS APIs do not offer.
  • Produce audio samples for language-learning content in under-served languages.

Try the demo

View model page

Read this issue on neuralexpedition.com →