Skip to content

Latest commit

 

History

History
101 lines (82 loc) · 6.67 KB

haas.md

File metadata and controls

101 lines (82 loc) · 6.67 KB

Haas-ja-v1.0

Haas-ja-v1.0 is a Japanese ASR model. It reuses Parakeet-TDT's FastConformer encoder and TDT joint, replacing the LSTM prediction network with two Kimi Delta Attention blocks. Use --family haas_asr to select this model.

Quick Start

./build/debug/bin/audiocpp_cli --task asr --family haas_asr \
  --model models/Haas-ja-v1.0-GGUF/haas-ja-v1.0-q8_0.gguf \
  --backend cuda --session-option haas_asr.matmul_weight_type=q8_0 \
  --audio japanese.wav --text-out transcript.txt --log

F16 and Q8_0 packages are available. The checkpoint is Japanese-only. The encoder uses full-context attention; the family's buffered streaming interface does not make it a native cache-aware streaming encoder. Long-form splitting uses the existing Parakeet runtime policy rather than upstream's quiet-point splitting policy.

The existing haas_asr.matmul_weight_type=q8_0 session option also keeps load-time packed and scaled projections in Q8_0. With native, projections constructed from float32 values retain float32 storage. Omit the override when using the F16 package if that precision is preferred.

Timestamp word boundaries follow the tokenizer’s word markers. Continuous Japanese text can therefore occupy one phrase-sized span; this does not add Japanese morphological word segmentation.

Model

Field Value
Family haas_asr
Tasks asr
Modes offline, streaming
Languages ja
Default package haas_ja_v1_q8_0
Input Japanese speech audio
Output Transcript text and optional timestamps

“Model/session default” means no explicit override. The runtime uses the checkpoint or session setting, or leaves the optional value unset as described.

Common Options (use directly)

Option Meaning
--model Model file or directory shown in the examples.
--backend Select an available backend supported by this model.
--threads CPU thread count.
--log Enable runtime logging.
--audio Input audio, where required by the selected task.

Request Options (use with --request-option)

Option Type / values Default Description
return_timestamps bool Model/session default Accepted for CLI --words-out compatibility. Word timestamps are always returned, including when false or omitted.
max_tokens int 0 Maximum TDT generated tokens; 0 or omitted uses the model-derived limit. Declared range: minimum 0.
keep_language_tags bool false Keep language tag tokens in decoded text; default false.
audio_chunk_mode auto, fixed, vad, none auto Offline audio chunking mode. fixed selects overlapping long-form windows regardless of offline_mode; vad uses the configured VAD model (Silero by default) to skip silence; auto follows offline_mode. Offline input is downmixed and resampled as needed.
audio_chunk_duration_sec float Model/session default VAD maximum chunk duration; omitted inherits the session duration. Fixed/long_form windows use the session setting, not this request override. Declared range: minimum 0.001.

Session Options (use with --session-option)

Option Type / values Default Description
haas_asr.weight_type native, f32, f16, bf16, q8_0 native Shared matmul weight storage type; default native.
haas_asr.matmul_weight_type native, f32, f16, bf16, q8_0 Model/session default Encoder and decoder matmul weight storage type; defaults to weight_type, which defaults to native. Q8_0 measured 1.79x faster on the tested CPU and changed roughly 8 percent of transcripts without moving aggregate word error rate.
haas_asr.conv_weight_type native, f32, f16 native Convolution weight storage type; default native.
haas_asr.perf_mode off, flash_attention off Encoder attention implementation. Default off uses the validated relative-attention path; flash_attention enables the fused implementation, which was numerically validated but slower on the tested hardware.
haas_asr.weight_context_mb int 3072 Weight context arena size in MiB; default 3072. Declared range: minimum 1.
haas_asr.encoder_graph_arena_mb int 1024 Encoder graph arena size in MiB; default 1024. Declared range: minimum 1.
haas_asr.decoder_graph_arena_mb int 256 Decoder graph arena size in MiB; default 256. Declared range: minimum 1.
haas_asr.audio_chunk_duration_sec float 2 Center-region duration for buffered streaming in seconds; default 2. Fixed context windows are re-encoded rather than cache-aware. Also sizes offline fixed/long_form windows. Declared range: minimum 0.001.
haas_asr.left_context_sec float 10 Past context included when re-encoding each buffered-streaming window in seconds; default 10. Also applies to offline fixed/long_form windows. Declared range: minimum 0.
haas_asr.right_context_sec float 2 Future lookahead included when re-encoding each buffered-streaming window in seconds; default 2 and adds equivalent partial-result latency. Also applies to offline fixed/long_form windows. Declared range: minimum 0.
haas_asr.streaming_attention_mode full_context full_context Attention policy inside each buffered window. full_context preserves bidirectional attention over the bounded window.
haas_asr.offline_mode full_context, long_form, auto full_context Offline encoder scheduling. full_context encodes the whole utterance, long_form uses bounded overlapping windows, and auto selects long_form beyond audio_chunk_threshold_sec.
haas_asr.audio_chunk_threshold_sec float 30.0 Duration threshold used by offline_mode=auto before switching to bounded-window long-form execution; default 30 seconds. Declared range: minimum 0.001.
haas_asr.vad_model_path path assets/framework/models/silero_vad VAD model path used by audio_chunk_mode=vad (any built-in VAD family); default assets/framework/models/silero_vad.

Convert Weights

To convert the official checkpoint, download Haas-ja-v1.0.nemo from the upstream repository, then run:

python tools/community_models/convert_haas.py \
  --source Haas-ja-v1.0.nemo --staging haas-converted \
  --converter build/debug/bin/audiocpp_gguf --type q8_0 \
  --output haas-ja-v1.0-q8_0.gguf

Conversion requires PyTorch, PyYAML, NumPy, safetensors, SentencePiece, tokenizers, protobuf and Transformers (tested with Transformers 5.12.1 and protobuf 5.29.5). NeMo and flash-linear-attention are needed only to run the official Python implementation, not for conversion or C++ inference.