TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #72

Generate live dialogue speech as the text streams in

Most open TTS models make you wait: they want the full sentence, then they render it. This one is built the other way around. It starts speaking the moment text begins arriving, which is exactly what a voice agent needs and exactly what most open models still can't do. It is small, permissively licensed, and has a demo you can talk to right now.

MODELGepard 1.0
PUBLISHEDJuly 15, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYAUDIO

Field notes

01What it does

Gepard is an autoregressive text-to-speech model built for real-time conversation. Instead of waiting for a complete sentence, it generates audio piece by piece as text arrives. On the vLLM serving path the first audio chunk lands in roughly 50 ms, fast enough that a dialogue feels live rather than played back.

The interesting part is how compact the recipe is. It is a single language model of about 556M parameters (a 14-layer Qwen3.5 backbone) that learned text and speech together, paired with NVIDIA's NanoCodec for audio. Because one model owns both modalities, the output carries natural rhythm and emphasis instead of the flat, stitched tone of classic TTS pipelines. On the public Seed-TTS-eval it leads the open field on perceived audio quality, with the highest naturalness score and the cleanest results on noise and discontinuity. In exchange for that streaming-first design it gives up some word accuracy and speaker similarity.

Voice cloning is zero-shot: a few seconds of reference audio captures the speaker up front, and cloning adds nothing to the per-word cost. It speaks English (US and UK), Mexican Spanish, Brazilian Portuguese, and Dutch.

02How to try it

Start with the demo Space. Type a line, pick a voice or upload a short reference clip, and listen to the stream come back. It is the fastest way to judge whether the prosody feels as natural as the benchmarks claim.

For local use, the weights are Apache-2.0 and the gepard-inference repo has the reference runner. If you care about the real-time numbers, use the gepard-vllm serving path; that is where the roughly 25x real-time throughput on a single RTX 5090 comes from. At about 556M parameters this fits on far more modest GPUs than most speech models.

03Caveat

Quality is strongest in English and varies by voice in the other languages. The streaming-first design also trades exact voice matching for speed, so if faithful speaker similarity is the priority, an offline model like VoxCPM2 still wins that column. Note the audio codec ships under NVIDIA's Open Model License rather than Apache, so check it if you are shipping commercially.

04What you can do with it

  • Build a voice agent that starts answering while the LLM is still finishing the sentence.
  • Clone your own voice from a short clip and narrate scripts or videos with it.
  • Serve many callers at once: one 96 GB GPU holds up to 256 parallel conversations.
  • Give a local assistant a natural voice in English, Spanish, Portuguese, or Dutch.
  • Batch-generate voiceover for content without paying a hosted-API meter.

Try the demo

View model page

Read this issue on neuralexpedition.com →

Gepard 1.0: generate live dialogue speech as the text streams in | Neural Expedition