MLX Audio 0.6.4 adds local speech synthesis, conversation and enhancement integrations

MLX Audio 0.6.4 adds integrations for local speech and audio work.

**Dek:** Neywa Labs’ November 2 release brings new audio models into the Apple-oriented inference runtime and fixes several issues that affected streaming output.

MLX Audio version 0.6.4 adds support for Audio8-TTS-Realistic 0.6B, Liquid AI’s LFM3-Audio 2B speech-to-speech model and NVIDIA NemotronLabs RE-USE speech enhancement. The November 2 release also changes several streaming paths, including fixes for Whisper text emission and for a timbre mismatch in the first chunk of a reference-conditioned decoder. The development is the runtime’s support for these workflows. The tagged release does not establish a new launch date for every model it integrates. [MLX Audio 0.6.4 release](https://github.com/Neywa-Labs/mlx-audio/releases/tag/v0.6.4)

Audio workloads need different checks

Speech synthesis turns prepared text into audio. Speech-to-speech can accept spoken input and return speech. Enhancement aims to make an input recording easier to use before or during later processing. A local application might need any one of those capabilities, or a combination. MLX Audio’s release gives developers more ways to run these tasks inside the same project, while each model still has its own requirements and behavior. [MLX Audio release](https://github.com/Neywa-Labs/mlx-audio/releases/tag/v0.6.4)

The streaming fixes illustrate why release verification should go beyond listening to a completed file. A recognizer can produce correct final text while a live interface briefly displays duplicated or broken characters. A voice decoder can produce a usable long clip while the first audio chunk has a different tone from the rest. Those problems matter in a conversation because the first partial result is often visible or audible before the final result arrives. The tagged notes identify the fixes, and testing a specific application still requires exercising that application's streaming integration. [MLX Audio release](https://github.com/Neywa-Labs/mlx-audio/releases/tag/v0.6.4)

A small local acceptance suite can expose the differences

Start with clean speech and a second recording with background noise. Measure whether enhancement improves the downstream transcript, then listen for any speech detail it removes. For the synthesizer, compare the first audible chunk with the rest of the output. For conversation, record how the client handles interruption and a delayed chunk. Keep the model revision and audio settings fixed when comparing results.

For teams choosing a local runtime to limit exposure of recordings, storage and diagnostics need their own review. A voice application may retain samples or partial transcripts even when the model runs on the same machine. Record which artifacts are needed to troubleshoot a failure and where they are stored. That gives developers an evaluation of their actual workflow, alongside the new integrations enumerated in the release.