Skip to content

Granite Speech TurboCTC MLX

Current status and review date

Decision snapshot

  • Product role: English STT 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: complete FP16 corpus quality; separate q8 static compatibility stop.
  • Decision: no-promotion; keep existing controls without newly qualifying them.
  • Important limitation: Critical entity and point-WER regression; q8 incompatibility is not an FP16 quality score.
  • Review dates: retained evidence and local documentation review through 2026-10-06 UTC.

Review narrative

Role: English STT challenger. Evidence: complete FP16 corpus; separate q8 static compatibility stop. Decision: no-promotion.

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

Measured FP16: iky1e/granite-speech-5.0-470m-turboctc-mlx-fp16@319abff7072204bc6cd30485aac58da6c1216214. Static q8 stop: iky1e/granite-speech-5.0-470m-turboctc-mlx-q8@f8911a51b3be9092a71fd0c676b9d8c3b9812035.

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

FP16 weights, no language-conditioning override, 16-kHz mono WAV. The same canonical human/synthetic, C1/C4 and holdout schedules are used. KV cache and LLM context are not applicable.

Exact reconstruction boundaries · Hardware view

Evidence by measurement class

Status: FP16 requests complete; q8 not loaded. Measured: canonical human 36/1,077 = 3.343% micro-WER, p50/p95 24.23/40.16 ms (72 warm C1 requests); holdout 82/1,864 = 4.399%. Limits: the critical MISSUS→miss error and point-WER regression fail gates; paired canonical interval −1.017 to +1.846 percentage points overlaps zero. Evidence: STT results and compatibility inspection.

Decision and promotion state

Fast component latency does not override critical or quality failures. No promotion. Static q8 incompatibility is not an FP16 quality score.

Failures and gotchas

The pinned q8 configuration declares global group size 128 while relevant quantized inputs use 64; the pinned loader lacks the needed override. No hidden conversion/runtime patch was admitted. Actual client acceptance and microphone/accent coverage remain missing.

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