Text to Speech on MacBook Air vs MacBook Pro

Short answer: For local text to speech on a MacBook, the laptop question comes down to memory, memory bandwidth, and thermal headroom rather than headline clock speed. A MacBook Air with 16GB of unified memory is enough for lightweight and mid-size models and for most narration work, while a MacBook Pro with more memory and more GPU cores finishes large-model batches faster and holds that pace longer under sustained load. If you generate short voiceovers, an Air is usually fine. If you clone voices across hour-long projects all day, the Pro's extra memory and cooling earn their price.

This guide is about running local text to speech on an Apple Silicon MacBook Air or MacBook Pro. It does not cover Intel Macs, Windows, phones, or cloud services. Everything below assumes a Mac running macOS 15 or later, which is the baseline for Murmur, the local voice studio used as the reference point here.

What actually decides TTS speed on a MacBook

Four chip properties govern how fast a local text to speech model produces audio on a Mac, and only one of them is the number people usually quote.

• Unified memory size decides whether a model fits at all. A local model has to be loaded into memory before it can speak, and if it does not fit, the system swaps to disk and the run crawls or fails.
• Memory bandwidth decides how quickly model weights move through the chip. Apple lists 153GB/s on the M5 MacBook Air, 307GB/s on M5 Pro, and 460GB/s or 614GB/s on M5 Max, so the same model can finish a long batch in noticeably different times across those tiers.
• GPU and CPU core count decide how much of the model runs in parallel per pass. More cores mean more work happens at once, which shortens each generation step.
• Thermal headroom decides whether that speed holds. A thin laptop can reach its limits and slow down during a long job, while a chassis with fans can sustain the same speed for hours.

Memory and bandwidth dominate for a specific reason. Apple's own MLX framework documents that Apple silicon uses a unified memory architecture where the CPU and GPU have direct access to the same memory pool, so arrays live in shared memory and operations on them do not need to copy data between a CPU and a separate GPU. You can read that explanation in the MLX unified memory documentation, and the framework's stated goal is efficient machine learning on Apple silicon. That is why memory size and bandwidth matter so much to a text to speech model: the weights sit in the pool the whole time it runs.

How much unified memory a local TTS model needs

In Murmur, the bundled Kokoro model is the lightest option and runs without a separate download. Its other models are optional downloads, and the app treats 16GB of unified memory as the practical floor for those models, steering a lower-memory Mac toward Kokoro instead. On Macs with 16GB or less, Murmur automatically selects quantized variants where it has them, trading a little precision for a smaller memory footprint. The table below shows the approximate download sizes the app lists in its model catalog, which is a good proxy for how much room each model wants.

Those figures are download footprints, not peak memory while running, which is higher once the model is loaded and generating. The practical reading is simple: on a 16GB Mac, the smaller models leave plenty of room for your other apps, the larger ones squeeze it, and the quantized variants exist exactly to close that gap. On an 8GB Mac, plan on the bundled model. If memory accounting is new to you, the local AI memory guide explains how unified memory is shared between the app, the model, and macOS.

MacBook Air vs MacBook Pro: a fit table

The current Apple Silicon lineup splits cleanly by memory bandwidth, and that split is the best quick predictor of how a laptop will handle local voice work. Use this as a fit guide, not a benchmark.

Two details matter when you read that table. First, the base 14 inch MacBook Pro with the M5 chip shares the Air's 153GB/s bandwidth, so it is not faster per step than an Air of the same generation; its advantage is cooling and a slightly larger battery. Second, the jump that changes local voice work is moving to M5 Pro, where bandwidth roughly doubles and memory options start at 24GB. That is the tier where large cloning models stop feeling cramped.

Battery and heat during a long local run

Apple's published battery figures are measured on video streaming and web browsing, so treat them as an upper bound, not a text to speech forecast. The 13 inch MacBook Air with M5 is rated up to 18 hours of video streaming and 15 hours of wireless web, and the 14 inch MacBook Pro with M5 up to 24 hours of video streaming and 16 hours of wireless web. Local generation is a different load: it keeps the GPU busy while the model produces audio, so the battery drains faster than those numbers suggest, and the gap widens as the model gets larger.

Heat behaves the same way. A MacBook Air has a thinner thermal budget, so a long batch can warm the chassis, and if the machine reaches its limits the chip throttles to stay safe. The MacBook Pro line has more room and active cooling, which is why it can hold a steady pace across a multi-hour queue. On either machine the practical advice is the same: plug in for long jobs, keep the laptop on a hard surface so air can move, and split huge projects into smaller batches rather than queuing one enormous job.

Run the light model when you are on battery and away from a desk, and save the large model for a plugged-in batch session.

Practical rule of thumb for local text to speech on a laptop

Set up a MacBook Air for TTS in six steps

1. Confirm your memory. Open System Settings, go to General, then About, and read the Memory line. That number, not the chip name, decides which models are comfortable.
2. Start with the bundled model. Kokoro needs no download and is the quickest way to hear whether local text to speech suits your workflow.
3. Match the model to your RAM. On 16GB, prefer the smaller models or the quantized variants the app offers; on 8GB, plan on the bundled model.
4. Close memory-hungry apps. Browsers with many tabs and video editors compete for the same unified memory pool.
5. Batch in chunks. Generate a few minutes at a time, listen, then continue, rather than queuing a very long job in one pass.
6. Run long queues on power. Plug in, and on an Air give the laptop airflow so it can shed heat.

To go deeper, the CPU vs GPU vs Neural Engine guide explains which part of the chip does which job during inference. The MLX text to speech overview covers the framework underneath, the Kokoro guide goes into the bundled model, and if memory is tight, 4-bit versus 8-bit quantization explains what the quantized variants change.

When a MacBook Pro is worth it, and when Murmur is not the answer

Reach for a MacBook Pro when the job is sustained and large: cloning a narrator's voice across a full audiobook, keeping several models loaded to compare takes, or running batches for hours without watching the machine. The extra memory and bandwidth also help when you run voice design alongside other apps. For occasional voiceover work, that difference is unlikely to justify the upgrade, and an Air will feel identical for short clips.

Murmur is a local voice studio for Apple Silicon Macs on macOS 15 or later, built for people who want the model on their own machine. It runs locally, so text and audio stay on the Mac and there are no per-character fees. It supports cloning from a reference sample, voice design, and multi-speaker projects, and it exports audio from a queue and timeline. It is a one-time $49 purchase with a free plan and a 3-day Pro trial, and a 7-day refund policy, so listen to the public samples before buying.

Murmur is also not the right tool for every job. It is macOS only, so it does not cover Windows, iPhone, Android, or a browser. It does not do live or real-time streaming text to speech. And it is a single-machine studio rather than a cloud collaboration platform. If you need any of those, choose a service built for them instead.

Frequently asked questions