OpenBMB family — efficient speech modeling; good entry if you follow Chinese OSS tooling around CPM.
Related searches & adjacent tasks
Readers often look for best TTS models, free TTS models, best speech recognition models, free speech recognition models, subtitle generation models, or free subtitle generation models. This page focuses on synthesis (TTS); recognition and subtitles usually rely on ASR (e.g. Whisper-class models) plus timing or translation—see the ASR hub link below.
Note: “Free” usually means open weights; commercial use still depends on each license and voice data policy.
More models & hosted API benchmarks
- Open-source list (task filter): https://huggingface.co/models?pipeline_tag=text-to-speech
- Speech recognition (ASR) models: https://huggingface.co/models?pipeline_tag=automatic-speech-recognition — useful for “best / free speech recognition” and as the first step of captioning.
- Hosted TTS APIs — quality, speed & price (third party): https://artificialanalysis.ai/text-to-speech/models?price=speed-vs-price — Artificial Analysis independent TTS comparison (e.g. Speech Arena–style quality metrics).
Selected Hugging Face repositories
Click a repo name to open the model card, examples, and license.
From k2-fsa (speech / streaming ecosystem). High engagement on HF—worth a look if you already use sherpa-onnx–style stacks.
MOSS team — tiny (~100M) TTS for teaching, ablations, and tight budgets; check the org for larger or realtime variants.
Mistral Voxtral (~4B) — aligns with Mistral licensing and API story; good if you want one vendor for LLM + speech.
Only 82M parameters but consistently top downloads—great for “get audio fast” prototypes and local narration; many community ports (ONNX, MLX, …).
Microsoft VibeVoice line—quality-oriented open weights; confirm exact parameter counts on the card.
Fish Audio S2 Pro (large capacity)—naturalness and multi-speaker / multilingual setups; often evaluated with Fish tooling and Spaces.
Smaller variant tuned for low-latency dialogue and streaming playback.
Qwen3 speech open weights—12 kHz line, custom voice / cloning focus per README; pair with Qwen LLM stacks for demos.
Disclaimer
Open weights ≠ unlimited commercial use. Check each License and voice-data statement. Cloud API pricing and SLAs are vendor-specific; Artificial Analysis helps compare quality, speed, and USD per 1M characters for hosted TTS.