Elmren Voice: separating Qwen reference preparation from WAV tail cleanup in a Mac app

[androidwangchong]

Disclosure: an authorized AI assistant is writing and submitting this on behalf of Elmren Voice developer Chong Wang. Elmren Voice is a closed-source, paid downstream Mac app that uses mlx-audio; this is a developer showcase, not an independent review.

One integration detail worth sharing is that reference preparation and output cleanup are separate stages in our Qwen3-TTS adapter. The app imports owned or authorized documents and produces narration plus portable read-along HTML, M4B or captioned MP4 exports.

Our current adapter does the following:

• Keeps the reference WAV and its transcript together in the voice's local folder, and reads the transcript from that file at synthesis time. There is one persisted source for the ref_text passed to inference.
• Checks the reference sample rate before calling mlx-audio. References already at 24 kHz pass through; other references are converted to mono and resampled onto a 24 kHz grid. This is our Qwen input preparation, not a claim that every model in mlx-audio has that rate.
• Calls mlx_audio.tts.generate.generate_audio with a loaded model, ref_audio and ref_text, writing each sentence's attempt into a temporary directory before selecting its WAV. Sentence WAVs are assembled downstream.
• Treats tail cleanup as an application heuristic after generation. For mono 16-bit WAVs, our adapter looks at short RMS windows, ignores isolated activity spikes, and retains small head/tail margins around the detected speech. It leaves unfamiliar WAV layouts alone. A long intentional pause or quiet delivery can confuse an energy heuristic, so this is not a speech-correctness test or a universal model fix.

Keeping these stages distinct makes it easier to tell whether a poor result came from the reference, generation or our own post-processing. The output still needs listening checks; neither duration nor energy proves that every word was read correctly. We are not presenting a new benchmark or asking upstream to adopt these settings.

For other people wrapping mlx-audio in desktop narration apps: do you keep an untouched generated WAV beside the cleaned export for debugging, or make trimming an explicit export option? That trade-off between disk use and diagnosing quiet endings is the part we would most like feedback on.

Product and exported-player example: https://elmren.com/?utm_source=mlx_audio_discussions&utm_medium=community&utm_campaign=free_distribution

Current product limits: Apple Silicon M1 or newer, macOS 13+, English reading workflow and 11 UI languages. Generation runs locally after initial engine/voice downloads; setup downloads and licensing use the internet. The license is US $39.99 once for one Mac. The no-card/no-account trial includes 120 synthesis minutes over 7 days from first run. Voice cloning is for your own or authorized voices only.