This video provides a comprehensive review and practical demonstration of Couti TTS (QITS), a newly released text-to-speech (TTS) model. The presenter details the model's claims of setting a new state-of-the-art in TTS performance, highlighting its low word error rate, high speaker similarity, and support for English and French. A key feature discussed is its "true streaming" capability, allowing for ultra-low-latency voice generation from large language models. While the model also features a speech-to-text component with voice activity detection, the focus is primarily on TTS. The video explores the model's restricted voice cloning policy, requiring anonymous voice donations. The presenter walks through the installation process, noting the absence of a simple web interface and demonstrating a custom Gradio app for testing. Performance tests on a mobile 4060 GPU show faster-than-real-time generation and modest VRAM usage. Various voice styles (default, happy, angry, disgusted, fast, laughing, whisper) are tested, showcasing the model's impressive emotional range, though some experimental results, particularly with "whisper" and "laughing" styles, are described as "creepy" or "uncomfortable." Overall, QITS is presented as a strong open-source competitor, especially given its MLX support for Apple devices and efficient performance.
Exploring the New Couti TTS Model: A Deep Dive into its Features and Performance
0:14 Introduction of Couti TTS (QITS): The video introduces a new text-to-speech (TTS) model, Couti TTS, which claims to set a new state-of-the-art in TTS performance with low word error rates and high speaker similarity.
0:44 Dual Functionality (TTS & STT): QITS includes both text-to-speech and speech-to-text functionalities, with the latter featuring built-in voice activity detection for automatic transcription.
1:24 Multi-Language Support: The model is noted for its ability to generate speech in both English and French.
2:04 True Streaming for Low Latency: A significant feature is its "true streaming" capability, allowing the model to process text as it's generated by an LLM, leading to ultra-low latency, crucial for conversational AI.
2:48 Restricted Voice Cloning: Unlike some models, QITS does not allow free voice cloning; instead, it offers the option for users to anonymously donate their voice to be cloned within a controlled environment.
3:50 MLX Support for Apple Devices: QITS supports MLX, enabling on-device inference for iPhone and Mac users, a notable advantage for open-source TTS models.
4:11 Installation and Setup: The installation process is described as relatively simple, requiring Python 3.12 and specific dependencies. However, the lack of a pre-built web interface led the presenter to create a custom Gradio app for demonstration.
9:23 Performance on Mobile GPU: Testing on a low-end mobile 4060 GPU showed impressive speed, generating 49 seconds of audio in approximately 41 seconds (faster than real-time) while using around 6GB of VRAM (including OBS recording).
7:36 Robust Voice Selection: The custom Gradio app allowed for testing a wide range of pre-defined voice styles including default, happy, angry, disgusted, fast, laughing, and whisper.
12:23 Emotional Expressiveness: The model demonstrated good performance with "angry" and "disgusted" emotional styles.
15:32 Unsettling Whisper and Laughter: Certain experimental outputs, particularly with "whisper" and attempting "laughter" using "haha" prompts, were described as "awful," "sensory uncomfortable," and "creepy," suggesting potential areas for refinement in emotional nuance.
17:09 Impressive Long-Context Handling: An extended test with a simulated office conversation showed the model handled "decently lengthy context" quite well, producing natural-sounding dialogue.
21:15 Overall Assessment: QITS is presented as a "worthy competitor" in the open-source TTS space, particularly due to its unique streaming feature and MLX support for on-device inference, despite some quirks in emotional expression.