mlboydaisuke/Whisper-large-v3-turbo-ExecuTorch

Whisper-large-v3-turbo — ExecuTorch (encoder + decoder)

Speech recognition in two .pte files: the encoder runs once per 30-second window, the decoder once per generated token. Putting them in one graph would re-encode the audio on every step.

graph	build	file	size (MB)	corr vs fp32 eager	ms	eager ms
encoder	XNNPACK fp32	whisper_large_v3_turbo_encoder_xnnpack_fp32.pte	2548.3	1.000000	1800.1	860.5
encoder	XNNPACK int8	whisper_large_v3_turbo_encoder_xnnpack_int8.pte	662.4	0.996407	1582.1	866.5
encoder	Core ML	whisper_large_v3_turbo_encoder_coreml_all.pte	1275.7	0.998309	589.1	869.7
decoder	XNNPACK fp32	whisper_large_v3_turbo_decoder_xnnpack_fp32.pte	953.3	1.000000	56.6	43.2
decoder	XNNPACK int8	whisper_large_v3_turbo_decoder_xnnpack_int8.pte	440.1	0.998471	46.7	40.7

Every file takes and returns fp32 tensors (token ids stay int64), so any encoder pairs with any decoder. The lightest working pair is 1102.5 MB.

• Source: openai/whisper-large-v3-turbo
• License: MIT
• Encoder input: log-mel spectrogram [1, 128, 3000] — 30 s at 16 kHz, 128 mel bins, hop 160, window 400, exactly what WhisperFeatureExtractor produces
• Decoder input: the encoder output plus decoder_input_ids [1, 128] int64, left-aligned and padded. Start with <|startoftranscript|>, a language token, <|transcribe|>, <|notimestamps|>.

Decoding

No KV cache: the decoder is a static graph over a fixed 128-token window, so a greedy step is take argmax of row len-1, append it, run again. Stop at <|endoftext|> (50257). 128 tokens covers a 30-second window of ordinary speech; past that, start a new window.

That costs a full 128-position forward pass per token, which is the price of a static graph that runs unchanged across runtimes and precisions.

Verification (Mac arm64, executorch 1.4.0, torch 2.13.0)

The two wrappers compose back to WhisperForConditionalGeneration exactly — max_abs_diff 0.000e+00 — and every graph matches torch fp32 eager at the correlations above. Timings are medians over 5 runs in one process: a relative reference, not a device number.

Two things worth knowing about the sizes

The decoder .pte is larger than the decoder's weights. Whisper ties proj_out.weight to decoder.embed_tokens.weight, but the two uses need different representations: an embedding table the portable kernels index into, and the same values packed into the XNNPACK delegate's blob for the output matmul. Tying them in PyTorch does not tie them here. Referencing the weight through F.linear instead of the proj_out module does not either — exported both ways, whisper-tiny's decoder comes out at 198.0 MB exactly.

The decoder's int8 build is the smallest portable one. Dynamic int8 quantizes the linear weights and leaves the token embedding table in fp32, and that table is 265.5 MB — 51,865 tokens at 1280 dimensions. There is no fp16 build at this size to compare against: int8 comes in at 440.1 MB and is the smallest portable decoder here.

Until recently there was no decoder int8 build at all, and this card said PT2E was observing the int64 decoder_input_ids. That was wrong on both halves. XNNPACKQuantizer.transform_for_annotation rewrites every scalar argument of add.Tensor/mul.Tensor as torch.tensor(float(arg)) whatever the node's dtype — one line in ExecuTorch's backends/xnnpack/quantizer/xnnpack_quantizer_utils.py, still present on main. In this decoder the casualty is position_ids = torch.arange(...) + past_key_values_length (modeling_whisper.py:749, past_key_values_length being a python int): it comes back float32, and the failure lands on self.weight[position_ids] — the position embedding lookup, not the token ids, and no observer involved. Measured by running prepare_pt2e with an empty quantizer and printing the failing node.

The decoder's Core ML build is not published. It converts, at 343.4 MB and correlation 0.961553 against fp32 eager, and then fails the only test that matters here: measured end to end — word error rate against the fp32 decoder: mean WER 43.9% (worst clip 66.7%) over 5 spoken sentences, the coreml_all decoder against the fp32 one with the other half and the waveform held identical; the fp32 arm transcribes all five correctly, so the comparison is against a working control rather than against noise.

Checked in the task's own units

Correlation is a first filter. These are the numbers that decide:

• encoder int8 — measured end to end — word error rate against the fp32 encoder: mean WER 0.0% (worst clip 0.0%) over 5 spoken sentences, the int8 encoder against the fp32 one with the other half and the waveform held identical; the fp32 arm transcribes all five correctly, so the comparison is against a working control rather than against noise.
• decoder int8 — measured end to end — word error rate against the fp32 decoder: mean WER 0.0% (worst clip 0.0%) over 5 spoken sentences, the int8 decoder against the fp32 one with the other half and the waveform held identical; the fp32 arm transcribes all five correctly, so the comparison is against a working control rather than against noise.

The sensitivity of that test, measured by injecting random noise into whisper-tiny's encoder output: rel_l2 0.03 (what int8 actually costs) and 0.10 both give WER 0.000; 0.20 and 0.40 give 0.025. Five clean sentences leave headroom, so a pass means does not break the transcript, not indistinguishable at any error level.

Conversion

python convert/export_whisper.py large_v3_turbo

The ExecuTorch tree ships a single-graph Whisper example under examples/models/whisper; this is that model with the halves separated.

(conversion scripts: executorch-models)