VoxCPM2 trades the per-character speech bill for a GPU and a base-URL change

Point an existing OpenAI audio client at `vllm serve openbmb/VoxCPM2 --omni` and the request body stays the same. The server exposes `/v1/audio/speech` and handles batching and KV-cache management itself, per the PyPI documentation the post cites ^[4]. The base-URL claim rests on a condition the post states outright: your application already calls OpenAI-shaped audio endpoints ^[5]. A client written against a vendor SDK needs new code in the client, not a new environment variable.

Throughput is where I would slow down. The project's PyPI page reports a real-time factor around 0.13 on an RTX 4090, against roughly 0.3 for the plain PyTorch path ^[6]. At 0.13, generation runs about 7.7 times faster than playback ^[2], and the PyTorch path runs about 2.3 times slower than that ^[3]. The 0.13 was measured on the Nano-vLLM-VoxCPM backend, while the production path the project recommends is vLLM-Omni ^[6][4]. The post reports the figure without batch size or input length ^[18]. For it to transfer you need that GPU class, similar text lengths and concurrency, and the same backend.

ElevenLabs' API quality tier costs $0.10 per 1,000 characters, according to the post ^[7]. That is $100 per million characters ^[1], or $1,000 a month at ten million ^[4]. Against that sits a box that loads the model in the 8 GB VRAM class before KV cache ^[12], plus the post's verdict that roughly 80% of routine generation can move and a hosted API stays for the hardest 20% ^[8]. The 80/20 is the author's judgement. The post's own advice is to run your sentences through the free Hugging Face playground first ^[15], which costs ten minutes instead of a deployment project.

One checkpoint covers plain TTS, voice design from a written description, and zero-shot cloning from a short reference clip; supply the reference transcript as well and pacing and emphasis carry over, not just timbre ^[9]. The design prompts the post gives are phrases like "hyped American on caffeine" ^[10]. The architecture is tokenizer-free: instead of quantising speech into discrete tokens, it generates continuous representations through a diffusion-autoregressive pipeline, which the post credits for better breathing and mid-sentence emotion ^[2]. Apache-2.0 means commercial use with no royalties or usage caps ^[16], and the post's second argument for self-hosting is that customer text stops leaving your network ^[17].

Install is `pip install voxcpm`, with weights pulled from Hugging Face on first run, driven either by the CLI or the Python API ^[11]. Lighter hardware can still run the VoxCPM-0.5B or VoxCPM1.5 checkpoints ^[13]. On Apple Silicon the community MLX-Audio path covers the 1.x checkpoints, and its documentation says VoxCPM2 support is not there yet ^[14].