First commit

.gitignore

@@ -0,0 +1,4 @@
+.env
+.venv
+**/__pycache__
+*.pyc

README.md

@@ -0,0 +1,22 @@
+# STT Runner
+
+Speech-to-Text transcription using sherpa-onnx + Qwen3-ASR.
+
+## Installation
+
+```bash
+python3 -m venv .venv
+.venv/bin/pip install -r requirements.txt
+```
+
+## Usage
+
+```bash
+python stt_runner.py \
+    --conv-frontend=path/conv_frontend.onnx \
+    --encoder=path/encoder.onnx \
+    --decoder=path/decoder.onnx \
+    --tokenizer=path/tokenizer \
+    audio1.wav audio2.wav ...
+```
+

core/args_parser.py

@@ -0,0 +1,14 @@
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--conv-frontend", type=str, required=True)
+parser.add_argument("--encoder", type=str, required=True)
+parser.add_argument("--decoder", type=str, required=True)
+parser.add_argument("--tokenizer", type=str, required=True)
+parser.add_argument("--language", type=str, default="", help="Force language, e.g. Indonesian, English, Chinese")
+parser.add_argument("--hotwords", type=str, default="", help="Comma-separated hotword phrases, e.g. 'foo,bar'")
+parser.add_argument("--num-threads", type=int, default=2)
+parser.add_argument("--provider", type=str, default="cpu", choices=["cpu", "cuda"])
+parser.add_argument("--max-total-len", type=int, default=2048)
+parser.add_argument("--max-new-tokens", type=int, default=256)
+parser.add_argument("sounds", nargs="+", help="Audio files to transcribe")

requirements.txt

@@ -0,0 +1,2 @@
+sherpa-onnx
+soundfile

stt_runner.py

@@ -0,0 +1,52 @@
+import sys
+from pathlib import Path
+import sherpa_onnx, soundfile as sf
+from core import args_parser as ap
+
+def stt_run(args):
+
+    print("Recognize...")
+    recognizer = sherpa_onnx.OfflineRecognizer.from_qwen3_asr( # qwen3 asr
+        conv_frontend   = args.conv_frontend,
+        encoder         = args.encoder,
+        decoder         = args.decoder,
+        tokenizer       = args.tokenizer,
+        hotwords        = args.hotwords,
+        num_threads     = args.num_threads,
+        sample_rate     = 16000,
+        feature_dim     = 128,
+        provider        = args.provider,
+        max_total_len   = args.max_total_len,
+        max_new_tokens  = args.max_new_tokens,
+    )
+    print("Recognizer ready!")
+
+    for f in args.sounds: # Multi-file
+        if not Path(f).is_file():
+            print(f"Skip. file not found: {f}", file=sys.stderr)
+            continue
+
+        audio, sr  = sf.read(f, dtype="float32", always_2d=True)
+        audio      = audio[:, 0]
+
+        stream     = recognizer.create_stream()
+        
+        if args.language:
+            stream.set_option("language", args.language)
+        stream.accept_waveform(sr, audio)
+
+        recognizer.decode_stream(stream) # Inference execution for `stream.result`
+        
+        text       = stream.result.text
+
+        if "<asr_text>" in text: # qwen3 asr format
+            text = text.split("<asr_text>", 1)[1]
+
+        print()
+        print(f"File      : {f}")
+        print(f"Duration  : {len(audio) / sr:.2f} s")
+        print(f"Result    : {text}")
+        print()
+
+if __name__ == "__main__":
+    stt_run( ap.parser.parse_args() )