Tortoise-TTS
The classic open-source multi-voice TTS: expressive prosody and reference-clip voice cloning from an autoregressive + diffusion design — famous for quality, infamous for speed.
About
Tortoise-TTS is the model that showed open-source TTS could sound genuinely human. Built by James Betker with the express priorities of strong multi-voice capability and realistic prosody, it chains an autoregressive decoder with a diffusion decoder — both notoriously slow samplers, which is why the author named it Tortoise and joked that a medium sentence takes two minutes on a K80. You clone a voice by handing it reference clips, and it picks up pitch, pace, and even quirks like lisps or stutters from them, along with the recording's ambience. It ships with quality presets (ultra_fast through high_quality), a simple do_tts.py / read.py CLI, a pip package, and a big community of forks that have since added streaming and cut latency well under half a second. The design is described in the paper 'Better speech synthesis through scaling' (arXiv 2305.07243). It remains a go-to for offline, no-API voice cloning, though the upstream repo has been quiet since late 2024 and every quality claim traces to the author's own listening tests.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported listening comparisons in the paper 'Better speech synthesis through scaling' (arXiv 2305.07243)
- No independent benchmark replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: peer-described design paper (arXiv 2305.07243); 14.9k GitHub stars; pip-installable with community forks keeping it alive. Strengths: Apache-2.0 licensed, reference-clip cloning with expressive prosody, fully offline, spawned a whole ecosystem of speed-up forks. Docked for: (1) upstream repo quiet since late 2024 — maintenance is now community-driven; (2) all quality evidence is author-reported listening tests, no independent benchmarks, and the stock pipeline is genuinely slow without the community streaming forks. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · James Betker's classic multi-voice TTS; repo created 2022-01-28, 14,886 stars, Apache-2.0
- · Autoregressive + diffusion decoder; reference-clip cloning picks up pitch/pace/quirks; quality presets ultra_fast to high_quality
- · Design paper 'Better speech synthesis through scaling' (arXiv 2305.07243) — all quality claims author-reported
- · Upstream quiet since late 2024; maintenance is now community forks (streaming, speed-ups); stock pipeline is genuinely slow