Introduction

Tontaube, a small speech-synthesis lab in Berlin, has released TontaubeV1.1. It's an open-weight text-to-speech (TTS) model with 2.9 billion parameters, built for zero-shot voice cloning, long-form narration and low-latency streaming. The weights went up on Hugging Face on November 12 (full model and 4-bit build), along with a matching v1.1.0 release of the inference server on GitHub. You can run it locally today, try it in the web playground, or call the hosted API.

What's new over TontaubeV1 (August): five style sliders for rate, pauses, intonation, pitch and articulation, an audio-refinement adapter, and a 4-bit version that's a 1.8 GB download and runs at about real time on a 12-core desktop CPU. No GPU needed. One catch up front: the weights are open, but the license isn't open source. It's free for research and for companies under EUR 200,000 in revenue.

What Shipped

  • Model: TontaubeV1.1 from Tontaube (Cremer & Cremer Technologies UG, Berlin). Two checkpoints: full bf16 (6.4 GB) and int4 (1.8 GB, the server's default).
  • Languages: English and German are the main ones. Spanish, French, Italian, Dutch and Portuguese also work, but Tontaube says native speakers haven't evaluated those yet.
  • Voice cloning: zero-shot, from about 5 to 60 seconds of clean, single-speaker audio. Bundled synthetic voices are included if you don't want to clone.
  • Speaking styles: three tags, audiobook, conversational and agentic.
  • Style sliders (new): rate (0–4), pause (0–4), pitch_range (0–5, monotone to very expressive), pitch_level (0–4, deep to high) and legato (0–2, crisp to normal). Leave one out and the model picks for itself.
  • Output: WAV files, or low-latency MP3 and Opus streaming. Long-form generation keeps a bounded context, so it's meant for chapters, not just sentences.
  • Under the hood: four stacked predictors. A Qwen3-1.7B-based model writes the semantic audio and timing, then three smaller Qwen3-0.6B-based models add acoustic detail. Audio is decoded with DualCodec and VibeVoice's acoustic decoder, plus a new small VibeVoice adapter that cleans up the result.
  • Retry safety net: a built-in recognizer checks English output and retries pieces where the model skipped text.

The Numbers

  • Speed on GPU (int4, RTX 5090): median 192 ms to first streamed audio, about 13× real time for one request and 21–23× across batched requests.
  • Speed on CPU: about real time on a 12-core desktop CPU with dual-channel DDR4. Long streams can pause between chunks.
  • Memory: the int4 build loads in about 4 GiB of GPU memory and peaks around 6.5 GiB. The full bf16 model on the vLLM backend needs an NVIDIA GPU with at least 24 GiB.
  • Quality (V1 only): Tontaube hasn't published V1.1 benchmarks yet. For V1, in an LLM-judged test of 400 English audiobook passages, it was level with ElevenLabs Flash v2.5 on prosody (50.1%) and preferred over Fish Audio S2 Pro (82.1%), Cartesia Sonic 3 (82.3%) and Gradium (86.2%). Tontaube notes that setup may favor its own model, because it used a cloned reference while some rivals used stock voices.
  • Hosted API: $5 per 1 million characters (about $0.0042 per minute of audio), with 200,000 free characters on sign-up, per Tontaube's pricing page.

Where to Run It Today

  • Local, CPU or small GPU (int4): clone the v1.1.0 server and start it. It downloads the int4 model on first run. Add LITE_DEVICE=cpu to force CPU. You need an x86-64 Linux CPU with AVX2, FMA and F16C, and a C++ compiler.

git clone --branch v1.1.0 --depth 1 https://github.com/craitech/tontaube.git
cd tontaube
uv run --python 3.12 --frozen tontaube serve

  • Generate a test clip from a second terminal once the server is up:

curl -s http://127.0.0.1:8080/predict \
  -H 'content-type: application/json' \
  -d '{"text":"Chapter one. The rain had not stopped for three days.","language":"english","tag":"audiobook","controls":{"rate":1,"pause":3},"format":"wav"}' \
| python3 -c 'import base64,json,sys; sys.stdout.buffer.write(base64.b64decode(json.load(sys.stdin)["audio_b64"]))' > chapter1.wav

  • Clone a voice: send a 5–60 second reference clip as voice_audio_b64 in the same request. The model card has a ready-made one-liner.
  • Full precision: start the server with TTS_BACKEND=vllm on a 24 GiB+ NVIDIA GPU to use the bf16 weights.
  • No setup: the playground and the pay-as-you-go API at tontaube.ai.
  • Heads-up: these aren't drop-in Transformers checkpoints. They need Tontaube's own inference server.

Caveats

  • Not open source. It ships under the Tontaube Community Model License 1.0. It's free for research, testing at any size, and commercial use while your company group makes under EUR 200,000 a year. Above that, you need a paid license, which you arrange by email for now. Commercial users must credit "Powered by TontaubeV1".
  • No reselling it as a TTS service. Hosting it so others can generate speech needs a commercial license at any revenue, paid or free. A small rate-limited demo is fine. Using it inside your own app falls under the normal revenue rule.
  • No V1.1 benchmarks yet. The only quality numbers are for V1, judged by an AI model, and English-only.
  • Very early. The repos had almost no downloads at release, and outside chatter is thin so far. Test it on your own scripts before you commit.
  • Usual autoregressive TTS risks. The card warns it can skip, repeat or change words and stop too early or too late, and German pronunciation isn't perfect yet.
  • Consent for cloning. Only clone voices you have permission to use.

Conclusion

TontaubeV1.1 is a strong option for narrators and indie developers who want expressive, cloneable English or German voices running on their own machine. You get real controls over pacing and pitch, and the 4-bit build plays back in real time on an ordinary desktop CPU. The license is the thing to read first: small studios and solo creators can use it commercially for free, but bigger companies and anyone hosting it as a service will need a deal. If you make audiobooks, game dialogue or video voiceovers, it's worth a test this weekend.

—Anabel ♡