Arm/whisper-large-v3-int8-litert

Whisper large-v3 optimized for Arm-based mobile CPUs with SME2

Whisper large-v3 automatic speech recognition optimized as an INT8 LiteRT .tflite model for Arm-based mobile CPUs with SME2.

Summary

This repository contains an Arm-optimized version of openai/whisper-large-v3 for automatic speech recognition. The model is provided in LiteRT .tflite, targeting Mobile CPU systems.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on LibriSpeech ASR and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	LiteRT .tflite
Target device class	Mobile CPU
Reference device	vivo X300 (C1-Ultra, C1-Premium, C1-Pro, Android 16 / OriginOS 6)
Primary performance result	14673.40 ms p50 end-to-end latency (RTFx 0.88x)
Accuracy result	WER 1.97% / CER 0.81%
Size / memory result	1493.66 MB, 3.94x smaller than FP32

Original model

Field	Value
Original model	openai/whisper-large-v3
Original source	Hugging Face
Original developer	OpenAI
Original model card	openai/whisper-large-v3
Original license	Apache-2.0

Model files

File	Description
whisper_large_v3_vivo_litert_optimized.tflite	Arm-optimized model for deployment
example.py	Minimal inference example
pyproject.toml	Pinned runtime dependencies for example.py, resolved with uv
uv.lock	Locked dependency resolution for pyproject.toml
config.yaml	Model I/O contract used by the example
benchmarks/	FP32 baseline and Arm-optimized benchmark records

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	vivo X300
CPU / accelerator	C1-Ultra, C1-Premium, C1-Pro (aarch64, 8 cores, cpu)
OS	Android 16 / OriginOS 6
Runtime	litert 0.9.0
Backend / delegate	XNNPACK + KleidiAI
Batch size	1
Precision	INT8 dynamic-range PTQ: per-channel symmetric INT8 weights, dynamically-quantized INT8 activations (no calibration data)
Runs	10 warmup + 50 measured

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 latency	342812.52 ms	14673.40 ms	23.36x faster
p90 latency	357440.97 ms	14834.61 ms	24.10x faster
Model size	5888.40 MB	1493.66 MB	3.94x smaller
Peak memory	8040.85 MB	5605.54 MB	1.43x less

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Field	Value
Dataset	LibriSpeech ASR (test-clean)
Split	test-clean
Number of samples	2620
Metric(s)	WER, CER
Evaluation runtime	litert

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
WER	2.03%	1.97%	-0.06 pp
CER	0.81%	0.81%	0.00 pp

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

Optimization area	Applied?	Notes
Model conversion	Yes	Converted to LiteRT .tflite via a multi-signature encode/decode export
Quantization	Yes	INT8 dynamic-range PTQ: per-channel symmetric INT8 weights, dynamically-quantized INT8 activations computed at runtime, no calibration data required
Runtime/backend selection	Yes	XNNPACK + KleidiAI, LiteRT runtime
Graph/runtime compatibility updates	Yes	Performed as part of the LiteRT export pipeline, including re-authoring the encoder and decoder into separate exportable modules with explicit self- and cross-attention KV cache tensors
Accuracy validation	Yes	Compared against the original model
Performance validation	Yes	Measured on the reference Arm platform

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

Install dependencies

Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:

uv python install
uv sync --frozen

Run the example

uv run example.py

Note: The Python/uv example runs on AWS Graviton (Ubuntu arm64) to confirm runtime compatibility only, and is intended as a guideline for building an equivalent run script on Mobile CPU systems.

Expected input

Property	Value
Input shape	[1, 128, 3000]
Input type	float32
Input range	N/A (log-mel spectrogram, not bounded to a fixed range)
Preprocessing	Load audio, downmix to mono and resample to 16000 Hz, then compute a 128-bin log-mel spectrogram via the Whisper feature extractor

Expected output

Property	Value
Output shape	Variable-length text string (decoded from up to 128 generated tokens)
Output type	string
Postprocessing	tokenizer.decode with skip_special_tokens=True, then strip surrounding whitespace

Intended use

This model is intended for developers evaluating automatic speech recognition workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• Accuracy was evaluated on the LibriSpeech ASR test-clean split (2620 samples) and may not generalize to other, noisier splits.
• Fixed maximum of 128 generated tokens per utterance.
• Greedy decoding only — no beam search, so transcriptions may differ from beam-search results reported elsewhere for Whisper large-v3.
• Evaluated only for English transcription (language forced to "en", task forced to "transcribe"); other languages and the translation task are untested despite large-v3 being a multilingual checkpoint.
• Latency and memory numbers are specific to the vivo X300 (aarch64, litert with XNNPACK/KleidiAI) — absolute numbers on other ARM chipsets, thread counts, or Android versions may differ.
• WER/CER are computed after Whisper-style text normalization, which affects the reported numeric values; raw, unnormalized transcripts will show different error rates.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.

Additional notes

Quantization uses a dynamic-range recipe: weights are INT8 per-channel and symmetric, quantized offline; activations are quantized dynamically to INT8 inside each kernel from an on-the-fly per-tensor scale, with INT8xINT8-to-INT32 GEMM in XNNPACK/KleidiAI kernels, then dequantized back to FP32 on the graph edges between kernels — no calibration dataset is required. The encoder and decoder were re-authored into separate exportable modules with explicit self- and cross-attention KV cache inputs/outputs, then converted into a single multi-signature .tflite file with "encode" and "decode" signatures for autoregressive decoding. The real-time factor (RTFx) improved from 0.04x on FP32 to 0.88x on the Arm-optimized model.

See example.py for the full greedy-decoding loop (prompt handling, token suppression, EOS stopping).

• Sample input: sample_input.flac is utterance 3575-170457-0005 of the LibriSpeech ASR corpus (test-clean split) by Panayotov et al., via OpenSLR (CC BY 4.0).

About this version

Original Model: openai/whisper-large-v3 by OpenAI - Repository

Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.

Converted/optimized by: Arm

License: The Original Model and the Optimized Model are subject to Apache-2.0.

This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.