Speech model families

A model family groups related recognition or speech-generation models. Choose both a model and a backend that can serve it, then select the matching Freehand profile from the table below. A different size, quantization, or server alias does not automatically change the profile; check the model guide’s requirements.

Want Freehand to install the model?

Use the managed runtime catalog. The families below also include models that require a service you manage yourself.

Families with a Freehand path

Family	Backend	Freehand model profile
Whisper, Distil-Whisper	Speaches, whisper.cpp, compatible transcription servers	Generic
Nemotron 3.5 streaming	NeMo-Speech.cpp	Nemotron
Parakeet TDT v3	NeMo-Speech.cpp	Parakeet TDT v3
Qwen3-ASR	vLLM	Qwen3-ASR
Cohere Transcribe	vLLM	Cohere Transcribe
Voxtral Mini 4B Realtime	vLLM	Voxtral Mini Realtime
Kokoro	Speaches, Kokoro-FastAPI	Generic
MagpieTTS Multilingual 357M v2602	NeMo-Speech.cpp	MagpieTTS
Qwen3-TTS 1.7B CustomVoice	vLLM-Omni	Qwen3-TTS
S1-mini	llama.cpp, vLLM, compatible chat servers	S1-mini

Family, runtime, or model format?

Whisper is a model family; faster-whisper and whisper.cpp are different implementations. GGUF, ONNX, and MLX describe packaging rather than new Freehand behavior profiles. A different size, quantization, or alias still needs a compatible backend and model profile. The server can run on any machine reachable from the Windows or macOS client.

Other families to consider

Exploration, not a compatibility list

These families have no dedicated Freehand model profile. Check the serving API before choosing a connection; a model download is not an API endpoint.

Recognition family	Distinct behavior to explore
Microsoft VibeVoice ASR Streaming	Streaming recognition with speaker-attributed output and hotword context. Its plugin server uses a different protocol from Freehand’s vLLM adapter.
NVIDIA Canary	Multilingual transcription and speech translation.
SenseVoice	Language identification, text normalization, and additional audio annotations in the FunASR ecosystem.
Kyutai STT	Streaming recognition with dedicated English and English/French variants.
IBM Granite Speech	A speech family with distinct checkpoint architectures, including English TurboCTC.
IndicConformer	Dedicated recognition across 22 Indian languages.

For speech generation, Chatterbox, Fish Audio S2, and OmniVoice offer different expression, voice, and language controls. XTTS, F5-TTS, CosyVoice, Piper, VoxCPM, and Higgs are also distinct families worth considering when choosing a speech server.

Some servers can already provide ordinary transcription or speech generation through a Generic connection if they accept its request fields and response formats. Generic does not enable a model’s specialized controls or an unrelated streaming API.

Reading model directories

Use Hugging Face’s recognition and speech-generation directories to discover families, then check:

• The task: general transcription, alignment, diarization, and language-specific derivatives can appear together.
• The serving API: popularity and download counts do not establish compatibility.
• The controls: choose the model and backend together for the workflow you need.