Hugging Face's logo

basyauma-tech
/
Haas-ja-v1.0-onnx

Automatic Speech Recognition
ONNX
Japanese
onnxruntime
cpu
japanese
speech-recognition
asr
parakeet
tdt
fastconformer
Model card Files
xet
 Community

Haas-ja-v1.0-onnx

English version below

非公式の変換物です。 third-intelligence/Haas-ja-v1.0のONNX版です。Third Intelligence, Inc.とは提携・承認・後援の関係にありません。

fp32は元のbf16をfloat32化。int8はONNX Runtimeの動的量子化版です。Hugging Face上の分類はquantizedですが、量子化はint8のみ。既定はfp32、int8は参考扱いです。int8はFLEURSの精度基準を満たしませんでした。

概要・環境

  • ONNX Runtime対応CPU向け。NeMo・CUDA・Triton不要。推論コード: haas-ja-local。
  • Python >=3.12,<3.13。実行時依存はonnxruntime・numpy・sentencepiece・soundfile・soxr。torch不要。
  • 計測はApple M5 Max（128GB）のCPUのみ。Superコア6・Performanceコア12、macOS 27.2（26B123）、Python 3.12.15、onnxruntime 1.31.1（CPUExecutionProvider、ORT_ENABLE_ALL）、numpy 2.5.4。
  • x86_64での速度・精度、Linux・Windowsは未計測。int8の精度は命令セットで変わる可能性あり。

使い方

hfはhuggingface_hubに付属（必要ならpip install huggingface_hub）。

pip install "haas-ja-local[onnx] @ git+https://github.com/basyauma-tech/haas-ja-local"
hf download basyauma-tech/Haas-ja-v1.0-onnx --local-dir Haas-ja-v1.0-onnx

from haas_ja_local import load_audio, load_model, transcribe

backend, tokenizer = load_model("Haas-ja-v1.0-onnx")
result = transcribe(backend, tokenizer, load_audio("audio.wav"))
print(result.text)
backend, tokenizer = load_model("Haas-ja-v1.0-onnx", variant="int8", threads=4)

haas-ja-local transcribe Haas-ja-v1.0-onnx audio.wav --backend onnx --variant fp32 --threads 4
haas-ja-local transcribe Haas-ja-v1.0-onnx audio.wav --backend onnx --variant int8 --threads 4
  • fp32/があれば既定はfp32、それ以外はint8。
  • 入力はsoundfile対応形式（WAV・FLACなど）。複数チャンネルを平均してモノラル化し、soxrで16kHzへ変換。
  • 30秒超は逐次分割。各断片の先頭から30秒±2秒内で、最も静かな0.1秒窓の中心を切断点とし、認識結果を連結。
  • --json --timestampsで1ファイル1行のJSONとトークン開始・終了時刻を出力。sherpa-onnxでの動作は未確認。

精度

asr-ja-bench v1.1、jacer v1.2.0。単位は%。FLEURSは650発話（2.36時間）、JVNVは1,375発話（3.33時間）。

版 サブセット CER strict CER flex CER kana 公表値strict 差（ポイント） 判定
fp32 FLEURS 3.69 3.02 2.38 3.65〜3.67 +0.02 合格
fp32 JVNV 4.52 3.06 2.25 4.50〜4.53 0.00 合格
int8 FLEURS 4.05 3.35 2.50 3.65〜3.67 +0.38 不合格
int8 JVNV 4.43 3.16 2.33 4.50〜4.53 -0.07 合格
  • 上記M5 MaxのCPU、バッチ1、8スレッド。発話ごとに認識し、サブセット全体でCERを計算（マイクロ平均）。句読点・空白を除外。全発話30秒以下で分割なし。
  • データはrebuild.pyで再構築し、書き起こしのハッシュを照合済み。JVNVの16kHz化は想定のffmpeg 7.1ではなく9.0.2のため、公式音声と微差の可能性あり。
  • 公表値は公式のNeMo・bf16・H200での測定。モデルカードと公開記事の差を幅で表示。差は近い端との差、範囲内は0。合格基準は公表値との差0.3ポイント以内。flex・kanaの公表値は未照合で、差・合否は判定していない。
  • 精度調整なし。int8でデコーダをfp32にしてもFLEURSのCER strictは4.09%で改善せず。配布int8はデコーダも量子化。
  • 未測定: JSUT・TEDxJP-10K・SpeechBSD（全サブセットの重み付きCERなし）。

詳細・再現手順: ONNXの検証記録、PyTorch参照実装の検証。

実装の確認

fp32とPyTorch参照実装（float32）はFLEURS全650発話で本文が一致。6発話の数値比較では最大誤差がlog-mel 9.54e-07、エンコーダ出力（相対）3.82e-06、結合ネットワーク出力（相対）1.11e-06で、トークン列はすべて一致しました。

PyTorch参照実装自体も第三者のfloat32 ONNXと6発話で比較済み。最大誤差はlog-mel 9.54e-07、エンコーダ（相対）4.36e-06、結合ネットワーク（相対）1.16e-06で、トークン列はすべて一致しました。

速度とメモリ

上記M5 MaxのCPU、バッチ1で逐次認識。FLEURSの先頭3発話をウォームアップとして除いた647発話、2.35時間。モデル・音声読み込みを除く認識時間で集計。GB＝10^9バイト。

版 スレッド数 RTFx中央値 RTFx合計 ピークRSS
fp32 4 72.4 72.7 3.24GB
fp32 12 66.7 66.4 3.24GB
int8 4 70.0 69.8 1.41GB
int8 12 84.3 84.3 1.40GB
  • RTFx合計は合計音声÷合計処理時間。ピークRSSはgetrusageのru_maxrss。
  • 12スレッドは全コアではなく、Super・Performanceコアにまたがる。fp32は4スレッドのほうが速く、再実行でも同じ傾向。原因は未調査。
  • CLIのfp32で33.8分（2,026.5秒）を68分割: 起動・モデル読み込み込み23.30秒（RTFx 87.0）、最大RSS 3.52GB。スレッド数は既定（ONNX Runtimeの自動設定）。
  • 公表RTFxはH200・バッチ512のため比較不可。4・12以外のスレッド数での速度は未計測。

ファイル一覧

合計サイズ: fp32約2.46GB、int8約0.62GB、リポジトリ全体約3.1GB。

サイズはバイト。SHA-256は各版のSHA256SUMSから転記。「—」は同ファイルに記載なし。

ファイル サイズ SHA-256
fp32/encoder.onnx 740,878 a93c1ac09be3a82cfd7d357affe4c84c2f4746a9f831f1ad04374fb744107adb
fp32/encoder.onnx.data 2,435,420,160 88689ebfe6a760d373598298b56716ca74ffee7f21e23d6a47a28a2557ad20f1
fp32/decoder_joint.onnx 27,712,720 9e1e99ec8d4a0ea43b19f8572b3fd02beae2dfb7f671e2ff975dd0d7428d626a
int8/encoder.onnx 3,198,708 f36c5a24e5a43c58321244b9b396f15f00277a48e93574a2bb79308c17cd3234
int8/encoder.onnx.data 609,822,464 03db20b51fd813f6b694309520621f66ea331dc7769f7679d936237a0abdb27a
int8/decoder_joint.onnx 7,137,058 ba4726a5a72f2673407435750f893d40c8cda04a34ac0a4a71c389802cf62c73
fp32/SHA256SUMS 248 —
int8/SHA256SUMS 248 —
config.json 640 —
tokenizer.model 279,711 —
preprocessor.npz 133,686 —
LICENSE 11,358 —
NOTICE 741 —

共通ファイルはモデル設定、SentencePieceトークナイザ、前処理の窓・メルフィルタバンク。両版のチェックサムは検証済み。

入出力仕様

opset 17。Bはバッチ、Tは入力フレーム数、T'はエンコード後のフレーム数。

ネットワーク 入出力 名前・形状・型
エンコーダ 入力 audio_signal [B,128,T] float32、length [B] int64
エンコーダ 出力 outputs [B,1024,T'] float32、encoded_lengths [B] int64
デコーダ＋結合 入力 encoder_outputs [B,1024,1] float32、targets [B,1] int32、target_length [B] int32、input_states_1/2 [2,B,20480] float32
デコーダ＋結合 出力 outputs [B,1,1,3078] float32（語彙3072＋blank＋duration 5）、prednet_lengths [B] int32、output_states_1/2 [2,B,20480] float32

input_states_1はKDAの再帰状態5×64×64。input_states_2はq/k/vのshort convバッファ3×320×4を先頭に置き、残りは0。

制限事項

  • 元モデル同様、フィラーが抜けやすく、漢字・かな・数字の表記指定は不可。無音・雑音で短語を出す可能性あり。
  • 30秒超の分割でCERが悪化。FLEURS 150発話を連結した33.8分音声のfp32のCER strictは、発話別4.10%、分割後6.39%。分割後のstrict/flex/kanaは6.39/5.70/4.62%でPyTorchと一致し、本文も同一。ONNXの数値誤差による悪化ではない。
  • 同一話者が話し続ける実際の長時間録音は未測定。
  • バッチ1のみ。ストリーミング・話者ラベル・公式Haas-ASRのカスタム語彙は未対応。
  • 時刻は80ms刻みのトークン単位。単語単位ではない。

ライセンスと帰属

English

This is an unofficial conversion of third-intelligence/Haas-ja-v1.0 to ONNX. It is not affiliated with, endorsed by, or sponsored by Third Intelligence, Inc.

fp32 casts the original bf16 to float32; int8 uses ONNX Runtime dynamic quantization. Hugging Face classifies this as quantized, but only int8 is quantized. fp32 is the default; int8 is for reference. int8 failed the FLEURS accuracy threshold.

Overview and requirements

  • CPU inference with ONNX Runtime; no NeMo, CUDA or Triton. Inference code: haas-ja-local.
  • Python >=3.12,<3.13; runtime dependencies: onnxruntime, numpy, sentencepiece, soundfile, soxr. No torch needed.
  • Measured only on the CPU of an Apple M5 Max (128 GB): 6 Super and 12 Performance cores, macOS 27.2 (26B123), Python 3.12.15, onnxruntime 1.31.1 (CPUExecutionProvider, ORT_ENABLE_ALL), numpy 2.5.4.
  • Speed and accuracy on x86_64, Linux and Windows are unmeasured. int8 accuracy may depend on the instruction set.

Usage

hf comes with huggingface_hub (pip install huggingface_hub if needed).

pip install "haas-ja-local[onnx] @ git+https://github.com/basyauma-tech/haas-ja-local"
hf download basyauma-tech/Haas-ja-v1.0-onnx --local-dir Haas-ja-v1.0-onnx

from haas_ja_local import load_audio, load_model, transcribe

backend, tokenizer = load_model("Haas-ja-v1.0-onnx")
result = transcribe(backend, tokenizer, load_audio("audio.wav"))
print(result.text)
backend, tokenizer = load_model("Haas-ja-v1.0-onnx", variant="int8", threads=4)

haas-ja-local transcribe Haas-ja-v1.0-onnx audio.wav --backend onnx --variant fp32 --threads 4
haas-ja-local transcribe Haas-ja-v1.0-onnx audio.wav --backend onnx --variant int8 --threads 4
  • Defaults to fp32 when fp32/ exists, otherwise int8.
  • Input: soundfile formats (WAV, FLAC, etc.). Channels are averaged to mono; soxr resamples to 16 kHz.
  • Audio above 30 seconds is split sequentially at the centre of the quietest 0.1-second window within 30 ± 2 seconds of each piece's start. Results are joined.
  • --json --timestamps emits one JSON line per file with token start/end times. Execution with sherpa-onnx is unverified.

Accuracy

asr-ja-bench v1.1, jacer v1.2.0; values in %. FLEURS: 650 utterances (2.36 hours); JVNV: 1,375 utterances (3.33 hours).

Variant Subset CER strict CER flex CER kana Published strict Difference (points) Result
fp32 FLEURS 3.69 3.02 2.38 3.65–3.67 +0.02 Pass
fp32 JVNV 4.52 3.06 2.25 4.50–4.53 0.00 Pass
int8 FLEURS 4.05 3.35 2.50 3.65–3.67 +0.38 Fail
int8 JVNV 4.43 3.16 2.33 4.50–4.53 -0.07 Pass
  • M5 Max CPU setup above, batch 1, 8 threads. Utterances transcribed individually; CER computed across each subset (micro-average), ignoring punctuation/spaces. All utterances ≤30 seconds; no splitting.
  • Data rebuilt with rebuild.py; transcript hashes verified. JVNV resampling to 16 kHz used ffmpeg 9.0.2 instead of the expected 7.1, so audio may differ slightly.
  • Published values use official NeMo, bf16 and H200. Model-card/blog differences form the range. Differences use the nearer endpoint, or 0 within the range. Acceptance threshold: within 0.3 points of published values. Published flex/kana values were not compared; no differences or pass/fail results are assigned to them.
  • No accuracy tuning. Keeping the int8 decoder in fp32 gave FLEURS CER strict 4.09%, with no improvement. The distributed int8 variant also quantizes the decoder.
  • Unmeasured: JSUT, TEDxJP-10K and SpeechBSD (no weighted CER across all subsets).

Details/reproduction (Japanese): ONNX verification, PyTorch reference verification.

Implementation verification

fp32 and the PyTorch reference (float32) produced identical text for all 650 FLEURS utterances. On 6 utterances, maximum errors were 9.54e-07 for log-mel, 3.82e-06 for encoder output (relative), and 1.11e-06 for joint output (relative); all token sequences matched.

The PyTorch reference itself was also compared against third-party float32 ONNX on 6 utterances: maximum errors were 9.54e-07 for log-mel, 4.36e-06 for encoder output (relative), and 1.16e-06 for joint output (relative); all token sequences matched.

Speed and memory

M5 Max CPU setup above, batch 1, sequential recognition. 647 FLEURS utterances, 2.35 hours, excluding the first 3 warm-ups. Recognition timing excludes model/audio loading. GB = 10^9 bytes.

Variant Threads Median RTFx Aggregate RTFx Peak RSS
fp32 4 72.4 72.7 3.24 GB
fp32 12 66.7 66.4 3.24 GB
int8 4 70.0 69.8 1.41 GB
int8 12 84.3 84.3 1.40 GB
  • Aggregate RTFx = total audio / total processing time. Peak RSS uses getrusage's ru_maxrss.
  • 12 threads span Super and Performance cores, not all cores. fp32 was faster with 4 threads, including on rerun; the cause is uninvestigated.
  • CLI fp32, 33.8 minutes (2,026.5 seconds), 68 pieces: 23.30 seconds including startup/model loading (RTFx 87.0), maximum RSS 3.52 GB; default thread count (ONNX Runtime automatic configuration).
  • Published RTFx uses H200, batch 512; not comparable. Speed with thread counts other than 4 and 12 is unmeasured.

Files

Total size: fp32 approximately 2.46 GB, int8 approximately 0.62 GB, entire repository approximately 3.1 GB.

Sizes in bytes. SHA-256 values are copied from each variant's SHA256SUMS; “—” means not listed there.

File Size SHA-256
fp32/encoder.onnx 740,878 a93c1ac09be3a82cfd7d357affe4c84c2f4746a9f831f1ad04374fb744107adb
fp32/encoder.onnx.data 2,435,420,160 88689ebfe6a760d373598298b56716ca74ffee7f21e23d6a47a28a2557ad20f1
fp32/decoder_joint.onnx 27,712,720 9e1e99ec8d4a0ea43b19f8572b3fd02beae2dfb7f671e2ff975dd0d7428d626a
int8/encoder.onnx 3,198,708 f36c5a24e5a43c58321244b9b396f15f00277a48e93574a2bb79308c17cd3234
int8/encoder.onnx.data 609,822,464 03db20b51fd813f6b694309520621f66ea331dc7769f7679d936237a0abdb27a
int8/decoder_joint.onnx 7,137,058 ba4726a5a72f2673407435750f893d40c8cda04a34ac0a4a71c389802cf62c73
fp32/SHA256SUMS 248 —
int8/SHA256SUMS 248 —
config.json 640 —
tokenizer.model 279,711 —
preprocessor.npz 133,686 —
LICENSE 11,358 —
NOTICE 741 —

Shared files provide model configuration, the SentencePiece tokenizer and preprocessing window/mel filterbank. Both checksum manifests have been verified.

Input/output specification

Opset 17. B is batch size, T is input frames, and T' is encoded frames.

Network Direction Names, shapes and types
Encoder Input audio_signal [B,128,T] float32, length [B] int64
Encoder Output outputs [B,1024,T'] float32, encoded_lengths [B] int64
Decoder + joint Input encoder_outputs [B,1024,1] float32, targets [B,1] int32, target_length [B] int32, input_states_1/2 [2,B,20480] float32
Decoder + joint Output outputs [B,1,1,3078] float32 (vocabulary 3072 + blank + duration 5), prednet_lengths [B] int32, output_states_1/2 [2,B,20480] float32

input_states_1 holds the KDA recurrent state, 5×64×64. input_states_2 holds q/k/v short conv buffers, 3×320×4, at the beginning; the remainder is 0.

Limitations

  • As with the original model, fillers may be omitted; kanji/kana/numeral style cannot be selected. Silence/noise may produce short words.
  • Splitting above 30 seconds raises CER. For 150 FLEURS utterances joined into 33.8 minutes, fp32 CER strict was 4.10% individually and 6.39% after splitting. Split strict/flex/kana scores were 6.39/5.70/4.62%, matching PyTorch; text was also identical. The degradation is not due to ONNX numerical error.
  • Real long recordings of one continuous speaker are unmeasured.
  • Batch 1 only. No streaming, speaker labels or official Haas-ASR custom vocabulary.
  • Token timestamps use 80 ms steps; they are not word timestamps.

License and attribution

Downloads last month
85
Inference Providers NEW
Automatic Speech Recognition
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for basyauma-tech/Haas-ja-v1.0-onnx

Base model

third-intelligence/Haas-ja-v1.0
Quantized
(14)
this model