csm vs Spark-TTS
Text-to-speech projects compared
csm and Spark-TTS both generate speech from text and are designed for local inference with GPU resources. csm focuses on speech conditioned by conversational context, while Spark-TTS adds bilingual synthesis and zero-shot voice cloning.
csm: Generate Conversational Speech from Text and Audio
CSM is Sesame’s speech-generation model, producing audio from text and optional conversation context. It suits developers building voice experiences who can run large models on a CUDA-compatible GPU and provide the required Hugging Face checkpoints.
Spark-TTS: Synthesize Speech with Qwen2.5
Spark-TTS is an efficient text-to-speech inference system built entirely on Qwen2.5, with zero-shot voice cloning and controllable speech generation. It supports Chinese and English and provides Linux installation instructions for Python 3.12+.
| csm | Spark-TTS | |
|---|---|---|
| Language | Python | Python |
| License | Apache-2.0 | Apache-2.0 |
| Stars | 14.7k | 11k |
| Forks | 1.5k | 1.1k |
| Last analyzed | Oct 3, 2026 | Oct 19, 2026 |
Key differences
- csm uses a Llama backbone with a Mimi audio decoder; Spark-TTS uses a Qwen2.5-based model to predict speech tokens and reconstruct audio.
- csm can use prior audio and transcript segments to shape conversational responses; Spark-TTS supports voice cloning from reference audio and adjustable virtual-speaker characteristics.
- Spark-TTS explicitly supports Chinese and English, cross-lingual cloning, and code-switching; csm’s README cautions that non-English performance is likely limited.
- csm generates audio rather than text and is intended to pair with a separate text-generating language model; Spark-TTS provides inference for text-to-speech and voice cloning.
- csm’s examples require access to CSM-1B and Llama-3.2-1B checkpoints on Hugging Face; Spark-TTS inference requires a separate 0.5B model, and its README documents Linux installation.
- Both projects use the Apache-2.0 license and Python; csm lists 14.7k stars and 1.5k forks, while Spark-TTS lists 11k stars and 1.1k forks.
Choose csm if you…
- need speech generation conditioned on prior conversational audio and transcripts.
- want a component that generates speech from text to pair with a separate language model.
- can run the required checkpoints on a CUDA-compatible GPU and follow its usage restrictions.
Choose Spark-TTS if you…
- need bilingual Chinese and English speech synthesis or cross-lingual voice cloning.
- want to clone a voice from reference audio or configure a virtual speaker’s gender, pitch, and speaking rate.
- prefer command-line inference or a Gradio UI, and can meet the project’s GPU and installation requirements.
Compare the numbers side by side, or add more projects
This comparison is generated with AI from the OSRepos analyses of both projects. Always check each project's repository and documentation before choosing.