Audio8 TTS on Apple Silicon: Setup and Honest Limits

Direct answer: Audio8 TTS Preview 0.6B has the clearest official Apple Silicon evidence among recent open TTS releases. The project publishes an INT4 ONNX CPU deployment and reports about 1 GiB of loaded memory on its tested Apple M2 setup. That path does not require CUDA, PyTorch, or Transformers after the model is downloaded. It supports local generation, streaming PCM, and reference-voice registration. Still, this is a preview checkpoint. The authors recommend eleven languages, warn that long or mismatched reference material can reduce stability, and advise short text segments for best quality. It is a strong experiment for technical Mac users, not proof of a finished creator app.

Quick decision table

What Audio8 actually ships

Audio8 TTS Preview uses a 601 million parameter DualAR speech architecture and includes its 44.1 kHz neural codec. The official model card lists Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, and Spanish as the recommended languages. Generation can work without a reference, while zero-shot cloning uses both reference audio and its exact transcript. That transcript requirement is important. A guessed or cleaned-up transcript can reduce similarity and stability even if the recording sounds good.

The model and code are released under Apache 2.0 with an upstream notice file. A permissive model license does not grant permission to clone a person. Obtain explicit consent, document the allowed project, and keep the reference private. The private voice cloning guide gives a practical consent and recording checklist.

Choose the ONNX route for a first Mac test

The main PyTorch instructions recommend a CUDA-capable GPU, so they are not the most direct Apple Silicon starting point. Use the repository's onnx_runtime guide and the official Audio8-TTS-Preview-0.6B-ONNX-INT4 package. That package uses weight-only INT4 for the Slow and Fast autoregressive models, FP16 activations and caches, and an FP16 codec. The project says normal synthesis loads only the generation and decoder sessions. During voice registration, it releases those sessions before loading the codec encoder to control peak memory.

Follow the pinned requirements in the official guide. Avoid mixing an arbitrary ONNX model with source from a different commit. Download into a dedicated directory, verify file sizes or checksums if provided, and run the smallest no-reference example first. Confirm that a WAV file is created and has the documented sample rate. Then start the local web or HTTP service only if you need it. Keeping the first test at the command line makes errors easier to isolate.

Register a reference voice carefully

Use a dry, single-speaker clip with no music, echo, or overlapping speech. The official card says the reference transcript must match the spoken content. Type every word, filler, and audible fragment accurately. Start with a neutral target sentence in the same language as the reference. Save the registered voice separately from the raw recording and document whose consent covers it. Do not publish reference audio merely to prove that cloning works.

Audio8 can accept long model contexts, but the project recommends keeping each input within 150 characters for best synthesis quality. Split narration on sentence or clause boundaries, generate numbered segments, and join them only after checking every boundary. For a production-oriented alternative, the Mac text-to-speech workflow explains how queueing, regeneration, and export affect a real project.

Measure the Mac you own

Do not turn the published M2 memory figure into a universal benchmark. Record your Mac model, unified memory, macOS release, ONNX Runtime version, model revision, thread settings, and whether the process is cold or warm. Measure time to first audio, total wall time, output duration, and peak memory. Run the same short, medium, and segmented-long scripts three times. An M1 with 8 GB and an M4 Max are both Apple Silicon, but their useful production envelopes are very different.

Quality needs its own test. Include names, abbreviations, dates, currency, quoted dialogue, and a final sentence. Mark missing words, repeats, wrong-language fragments, clipped endings, and unexpected silence before rating naturalness. Compare at matched loudness with file labels hidden. The narration voice guide offers criteria that are more useful than choosing a demo by first impression.

Limitations and production risks

• Preview status means language and dialect coverage can change.
• The published 1 GiB figure is from one tested Apple M2 configuration, not every Mac.
• The recommended text length is short enough that narration needs segmentation.
• Cloning quality depends on clean audio and an exact reference transcript.
• The ONNX package reduces runtime dependencies but does not provide project management or editorial review.
• Generated voices can be misused, so consent and disclosure remain part of the workflow.

Use or skip verdict

Use Audio8 on Apple Silicon if you want to evaluate a compact multilingual clone model through an official CPU path and you are comfortable assembling the workflow. It is especially attractive for developers who need an embeddable local service. Skip it when your priority is opening a Mac app, managing many scripts, recovering failed jobs, and exporting consistently. Murmur costs $49 one-time, has no free trial, and includes a 7-day refund policy. Compare the complete workflow, not only model weights. The AI voice generator for Mac guide helps frame that decision.

Practical checklist

1. Confirm you are using the official Audio8 repository and ONNX INT4 package.
2. Pin the repository revision, model revision, and ONNX Runtime version.
3. Run no-reference synthesis before registering a voice.
4. Use a consented clean clip and an exact transcript for cloning.
5. Keep inputs near the project's recommended short length and segment narration.
6. Measure warm speed, first audio, peak memory, completion, and silence on your Mac.
7. Retain raw outputs and rerun the same script after upgrades.

Frequently asked questions