> Source: GitHub timoncool/HiggsAudio-Studio + file-share package + note "6GB VRAM" | Six verification channels: repo API + full README + HF metadata (v3/v2) + boson.ai official page + LMSYS blog
Conclusion first: Studio is a community developer's portable TTS tool (84 stars, created 06-13, still updated yesterday). Its upstream is Boson AI's Higgs TTS 3—4.65B parameters, released 06-04, 100K downloads, Research & Non-Commercial license. The noted "6GB VRAM" has no official figure in the README; it traces to Studio's online quantization mechanism ("auto by VRAM," experimental): 4.65B fp16 requires 9.3GB, so 6GB roughly matches real-world int8 online quantization. Reasonable—but it's a resource-pack measurement, not an official threshold. Usable with source attribution.
The Tool's Function Is One Pipeline
TTS (100+ languages) + zero-shot cloning (reference audio auto-transcribed by Moonshine ASR) + an AI text director + podcast mode + audiobook mode, fully offline, with Python, dependencies, models, and caches all in one folder—a "portable/green" philosophy where deleting the folder uninstalls it.
The director is the most interesting part: a local GGUF model (Qwen3.5-9B, 5.5GB) rewrites plain text into text with control tags—43 tags: 21 emotion, 10 prosody, 3 style, 9 sound effects—which then go to Higgs TTS 3 for synthesis. Podcast mode adds loudness alignment (LUFS −16, the podcast industry standard) with a distinct voice per speaker.
The pipeline splits into three layers: the upstream model handles acoustics, the director LLM handles annotation, and the tag vocabulary is the contract between them. Voice "performance instructions" become an independent interface layer here—a different kind of interface observation from the 09-01 Luna-TTS post (which was about decode order and frame rate); this one is the last representation layer before text becomes speech.
Against VibeVoice: Two Voice Worlds in One Night
Yesterday's post covered VibeVoice, today Higgs—a natural controlled comparison. VibeVoice 7B, 14GB VRAM, official Microsoft, TTS code retracted with a pivot to ASR; Higgs 4B-class, 6GB-tier VRAM, a startup pushing full speed—v3 not only wasn't retracted, emotional control became an official feature (inline control tags placed directly in input text, integrated same-day by SGLang-Omni on 06-04).
Generation-side risk control thus takes two forms: Microsoft chose retraction, Boson chose licensing—Research & Non-Commercial, with commerce routed through the Boson Workspace paywall. One closes the gate, one builds a toll booth; neither leaves a door open for abuse.
On the community side, Studio wraps the upstream tag vocabulary into engineering (button-inserted tags, director auto-annotation, batch synthesis)—an 84-star project turning a paper feature into a production line, the same move Ollama made for open-source LLMs.
A deployment footnote: in Studio's platform support table, Linux aarch64 lists DGX Spark / Jetson (CUDA 13.0). In the edge-side lineage, Spark has already been hit three times by TensorFold, Thor-U, and Strata—this voice tool's support table is the fourth. Spark as an inference target is spreading from review posts to ecosystem defaults.
Two falsifiable predictions:
1. Within 12 months, a cross-model de facto standard for speech emotion control tags emerges (the closed 43-tag vocabulary is a ready starting point; the HF community will converge on a convention). 2. Research NC-licensed upstream models will produce a two-tier structure: "flourishing non-commercial toolchains, commerce via official API"—projects like Studio stay on the non-commercial side long-term.
Finally, the minimal path: a 6GB VRAM card runs Studio's core TTS line; 8GB+ is recommended to also enable the director model. When director and acoustic models fight over VRAM, protect the acoustics first—the director can swap down to the 2.5GB Qwen3.5-4B.
---
*Verification chain: GitHub API (stars/timeline/no license file = inherits upstream terms), full README (43-tag list, auto by VRAM, DGX Spark/Jetson support table verbatim), HF API (bosonai/higgs-tts-3-4b: created 06-04, 101,308 downloads, 4.65B params), boson.ai official page ("speaks, not just reads"), LMSYS/SGLang-Omni blog (inline control tags integrated same day). The 6GB VRAM figure is a resource-pack measurement; it does not appear in the README and is flagged as such.*