Downloads · 30 days
9
11% of all-time downloads
miguelamendez/mica-granite-speech-5
mica-granite-speech-5 is a automatic speech recognition model from miguelamendez. Use it when you need speech turned into text. It is set up for transformers. The card lists the license as apache-2.0.
Mica runtime artifacts for ibm-granite/granite-speech-5.0-470m-turboctc, pinned to revision 6c14d3d052a602d850f1bc4aa017f25f4adf6aa0. The selected checkpoint is Apache-2.0; it is not the separate non-commercial -nc ch…
Downloads · 30 days
9
11% of all-time downloads
All-time downloads
85
Public
Repo size
1.6 GB
Likes
0
Public
Click a slice to open those files.
.safetensors856 MB · 52%
From the Hugging Face model README
Mica runtime artifacts for
ibm-granite/granite-speech-5.0-470m-turboctc,
pinned to revision 6c14d3d052a602d850f1bc4aa017f25f4adf6aa0.
The selected checkpoint is Apache-2.0; it is not the separate non-commercial
-nc checkpoint.
The runtime configuration, conversion policy, and validation evidence are
maintained in the Mica Server repository.
Validated artifacts are published together in the
miguelamendez/mica-granite-speech-5
repository, with manifests recording their source, size, and digest.
Granite Speech is a 470M-parameter English CTC ASR model. It has 16 conformer blocks, 1,024 hidden dimensions, eight attention heads, 128-frame block attention, temporal subsampling from 100 Hz to 12.5 Hz, and a 16,384-unit BPE output head. Greedy CTC decoding is non-autoregressive.
Upstream reports approximately 60,000 hours of English training audio from public datasets plus disclosed synthetic mixtures. It does not publish a single maximum training utterance duration or supported input-audio duration. The encoder configuration's 128-frame attention block and 512 positional field are architecture internals, not a claim that audio must end at either value.
Token context, generated-output limits, KV-cache precision, and thinking levels are therefore not applicable. Granite profiles use maximum audio duration, chunk duration, overlap, batch size, and measured workspace instead.
mlx/q4/ selective MLX affine Q4, group size 64
mlx/q8/ selective MLX affine Q8, group size 64
gguf/ Q4_K and Q8_0
vllm/ not published; structural candidates failed the production gate
The MLX conversion retains these modules at source precision:
encoder.input_linear, the sole learned acoustic-feature input boundary;encoder.out, shared by midpoint and final CTC vocabulary logits;encoder.out_mid, which feeds midpoint CTC predictions back into the encoder.ctc_head, the tied public output head required by the Transformers loader.Uniform Q4 corrupted the controlled transcript. Selective Q4 and Q8 matched the BF16 transcript exactly on the fixed 8.608-second quality clip.
GPTQ W4A16 and W8A16 group-128 structural candidates now calibrate from real
audio manifests. Both reload and match the BF16 fixture transcript in batch-2
Transformers inference. They remain unpublishable: the calibration set has one
record, and vLLM core still has no native granite_speech5_ctc transcription
loader.
On an Apple M4, selective MLX Q4 and Q8 passed real HTTP transcription. Five warm runs after warmup had median model times of 0.0637 and 0.0635 seconds for a 3.505-second clip, about 135x real time. GGUF Q4_K and Q8_0 also passed real Metal inference. These short-clip results do not certify unlimited streaming duration, multi-hour files, or concurrent batching.
Validation records:
docs/validation/granite-speech-5-mlx-q4.mddocs/validation/granite-speech-5-mlx-q8.mddocs/validation/gguf-runtime-macos-metal.mddocs/validation/granite-vllm-quantization-macos.mdA Granite profile must define max_audio_seconds, chunk_seconds, overlap,
maximum concurrent streams, and a certified memory peak. Values beyond the
longest Mica-tested clip warn in experimental mode and fail certified mode.