SpeedASR-v1-65M
SpeedASR-v1-65M is an offline English Zipformer transducer with 65,551,059 parameters (65.55M). This release contains the original FP32 checkpoint, its 500-entry token vocabulary, portable batched inference code, and public Open ASR Leaderboard evaluation artifacts.
Original encoder/decoder/joiner ONNX exports are also retained in onnx.
Their epoch-99 filenames are export labels around the epoch-30-avg-9 checkpoint
alias, not a claim that this is an epoch-99 model. The reported scores and cloud
adapter use PyTorch; these ONNX files have not been evaluated for numerical
parity by this release's checks.
The non-causal encoder processes complete recordings. This release does not provide a native streaming recognizer or claim a measured phone/NPU benchmark.
| Property | Value |
|---|---|
| Language | English |
| Encoder | Zipformer2, six stacks / 16 encoder layers; non-causal |
| Prediction network | Transducer, context size 3, dimension 512 |
| Joiner | Dimension 512 |
| Parameters | 65,551,059 total |
| Audio | 16 kHz, mono, floating-point samples |
| Features | 80-bin kaldifeat Fbank, dither 0, snip_edges=False, high_freq=-400 |
| Decoding | Modified beam search, beam size 4, FP32 |
| Text formatting | VoiceA casing/punctuation tags restored; no benchmark WER normalization inside inference |
| External language model | None |
| Checkpoint | pretrained.pt, 262,613,934 bytes; epoch-30-avg-9 |
The complete frozen construction and inference settings are in model_config.json. The repository's existing Apache 2.0 license is preserved. Vendored Icefall components retain their upstream notices and Apache 2.0 license.
Inference
The pinned dependency set targets Linux x86-64, Python 3.10, CUDA 12.4 and PyTorch 2.6.0. The k2 and kaldifeat wheels must match the PyTorch/CUDA ABI.
python3.10 -m venv .venv
source .venv/bin/activate
python -m pip install huggingface_hub==1.18.0
hf download SpeedAudio261119/SpeedASR-v1-65M --local-dir SpeedASR-v1-65M
cd SpeedASR-v1-65M
python -m pip install -r requirements-linux-cu124.txt
python inference.py /path/to/16khz.wav --device cuda:0
From a downloaded copy of this repository:
import soundfile as sf
from inference import SpeedASR
audio, sample_rate = sf.read("sample.wav", dtype="float32")
model = SpeedASR(device="cuda:0")
text = model.transcribe(audio, sample_rate=sample_rate)
texts = model.transcribe_batch([audio, audio], sample_rate=sample_rate)
print(text)
Inputs must be mono and 16 kHz; resample other audio before calling the API. The WAV CLI uses the first channel, matching the benchmark frontend. Inputs shorter than one second are padded; inputs longer than 30 seconds are decoded in consecutive 30-second pieces and joined without overlap. The same policy and decoder settings apply to all datasets.
The loader checks the checkpoint/token hashes, loads the state dictionary
strictly, and verifies the parameter count. SpeedASR.from_pretrained can
download weights, tokens and configuration for an explicitly pinned Hub commit;
the local inference module and speedasr_core must already be available.
Public benchmark results
The following local measurements were completed on 19 November 2026
using an NVIDIA GeForce RTX 3090 Ti, FP32, with the official scorer at
4a113dd90c9ffd011d7d3d105da76980d511d029.
They are not official H200/cloud results or an official leaderboard ranking.
| Public testset | Clips | WER (%) |
|---|---|---|
| AMI Cleaned | 7,805 | 9.59 |
| GigaSpeech Cleaned | 18,768 | 8.07 |
| VoxPopuli Cleaned AA | 628 | 3.27 |
| Earnings22 Cleaned AA Chunked | 341 | 6.78 |
| SPGISpeech | 39,341 | 3.89 |
| Voice Arena Monsoon English | 2,102 | 4.44 |
| URGENT2024 | 1,000 | 12.29 |
| Seven-set arithmetic mean | 69,985 | 6.9043 |
| URGENT2024 Clean β optional, excluded from mean | 1,000 | 5.54 |
Earnings22 predictions are reconstructed and scored as six parent calls, not 341 independent references. The two private score groups are unavailable locally and are excluded from this public average.
Seven-set pooled local throughput was 192.6293 RTFx, using up to 2,048 streams, a 5,000 padded-audio-second budget, and 512 long-audio chunks per internal batch. This is batched GPU throughput, not single-request latency or on-device efficiency. Cloud throughput must be measured separately.
The benchmark_results/local_3090ti_20261119 directory includes all 70,985 raw public prediction rows, final scores, dataset revisions, sanitized historical configuration, and verification data. Only machine-local audio paths and model-id filenames were changed for release; raw references/predictions, timing and chunk metadata were preserved. Dedicated Earnings22 and Monsoon filenames use the official empty-configuration convention. The official public summary was generated directly from these released local outputs. export_manifest.json records original and released hashes. It contains no audio corpus or private testset data.
The portable runtime passed strict loading and exact prediction comparisons against the original runner on 24 representative short/middle/long examples covering all eight public sets, plus a 16-example identical-batch comparison. See portable_validation.json. This is a targeted portability check, not a rerun of all testsets with the new cloud adapter. Different batch grouping can cause numerical differences.
The adapter smoke check also passed
official data preparation, raw manifest writing and scoring using four cached
AMI utterances and one complete 51-chunk Earnings22 parent call. Dataset transport
and metadata requests were replaced with local test fixtures for that check.
Docker was unavailable on the validation host, so a clean image build and an
actual H200 Job remain subsequent validation steps.
The combined pinned inference/evaluation dependencies passed a fresh
pip --dry-run --ignore-installed resolution; see
dependency_validation.json.
Open ASR cloud evaluation
leaderboard/run_eval.py uses the official data loader and manifest writer, retains raw references and predictions, preserves Earnings22 parent/chunk metadata, and uses one scoring worker. It fixes the decoder settings above across datasets and records synchronized inference timing after synthetic warmup.
The Dockerfile, HF Jobs launcher, and submission instructions prepare a Docker Space and the seven public cloud evaluation jobs. Build with:
docker build -f leaderboard/Dockerfile -t speedasr-open-asr .
Use an immutable model commit SHA for submitted runs. H200 batch settings should be checked for the maximum practical throughput; the included defaults reproduce the local limits and are not claimed to saturate an H200. Uploading this package does not itself build a Space, launch paid Jobs, submit a GitHub PR, evaluate private sets or register an official leaderboard entry.
Training data
The author confirms that all seven default public benchmark testsets listed above were excluded from training. This is an author-provided statement, not an independent overlap audit. The full training-dataset composition has not been disclosed in this release. No exclusion claim is made here for unavailable private datasets or the optional eighth testset.
Artifact identity
pretrained.pt SHA-256
b0475d24e50f810fbcf997e66061243a1fe212251f6d032457b2711641906c01
tokens.txt SHA-256
a238b7a0fa67aa6be998ca13dd129d69c50639970583e32232dd149e5448210a
runtime_source_manifest.json records vendored
source origins and transformations. The upstream model-code revision is
3f848bb6d0acc970c9b294a30ca0a04a7c9c78d1. The release artifact manifest
records the files actually published.
leaderboard/model_metadata.json provides structured submission metadata and parameter counts by component, including auxiliary training projections retained in the unchanged checkpoint.