Skip to content

Breeze TTS 2 MLX

Current status and review date

Decision snapshot

  • Product role: English native-voice TTS challenger.
  • Selected or best-qualified configuration: none qualified; exact tested configurations are documented below.
  • Measured hardware: Apple M4 Max, 40-core GPU, 48 GiB unified memory; direct same-host component HTTP.
  • Evidence: BF16 resource failure; full q8 component retries and independent ASR proxy.
  • Decision: no-promotion; keep existing controls without newly qualifying them.
  • Important limitation: No human naturalness preference or actual playback acceptance; acoustic ambiguities unresolved.
  • Review dates: retained evidence and local documentation review through 2026-10-06 UTC.

Review narrative

Role: English native-voice TTS challenger. Evidence: BF16 resource failure; complete q8 streaming and buffered retries. Decision: no-promotion; no human naturalness winner. The model family is not terminally rejected.

Evidence reviewed: 2026-10-06 UTC. This is a dated snapshot for an English personal, noncommercial, quality-first assistant, not current live-service reporting.

Immutable identity

BF16: mlx-community/Breeze-TTS-2-mlx@3c8829fb7fd335818f085cd2ef49b4100c0e46c8. q8: mlx-community/Breeze-TTS-2-mlx-8bit@c6e4a2ff6ab9afba68b7853de802273ffe23fb49.

Candidate audio runtime: MLX Audio 0.5.8, source 70f4add32911bab6f869b824864ad9f1e24dcb97, MLX 0.32.3. The reviewed Anvil audio launcher file SHA is e7149c09668713835e1cd70725f2af78608ac023c72541f96462c61d7c924661; this file hash is not a Git commit. Runtime distribution/source identity and exact local snapshot gates are retained. License/source screening is recorded separately from local qualification in the source registry.

Tested hardware and topology

Apple M4 Max, 40-core GPU, 48 GiB unified memory. These are direct same-host component HTTP measurements. Private hostnames, addresses, operator identity and paths are redacted. Historical CUDA or other-device evidence is a separate population.

Engine, quantization, KV, context, and concurrency recipe

Native voice S0; 24-kHz source normalized to 16-kHz PCM. BF16 uses streaming with the 1,200-token HTTP default; q8 explicitly uses 750 tokens, cache zero and worker memory trace. R2 streams; an eight-case diagnostic then full R3 changes only request streaming to false. C1, 60 distinct prompts plus one separate cold request. No KV override.

Exact reconstruction boundaries · Hardware view

Evidence by measurement class

Status: BF16 59/60 before managed resource abort; R2 and R3 60/60 warm. Measured: R2 first PCM p50/p95 1,527.19/1,582.73 ms; buffered R3 5,458.13/17,346.36 ms. R3 ASR proxy micro-WER 7.749%; 435 guard samples, zero observed swap growth, minimum free memory 42%. Limits: no human votes; possible unit omission, added “Yeah,” name pronunciation and pauses require acoustic review. Forty-one digital-zero repeat flags are not confirmed spoken loops. Evidence: TTS results and final assessment.

Decision and promotion state

The exact tested configurations remain unqualified. Keep the existing Kokoro control without calling it best naturalness. Full completion/resource scope cannot replace blind preference and actual client gates.

Failures and gotchas

Cache zero controls free buffers, not active memory. The BF16 swap-growth cause is not established; retry residual swap differs. Buffered first-audio timing does not establish interactive playback. Lower ASR WER is not an acoustic winner. One private-name case is redacted with original hashes and metrics retained.

Public artifacts are declared sanitized derivatives, with original private SHA/bytes in the projection ledger. Completion or documentation does not authorize a service, route, or client-catalog change.

Dated run history