Integrate CrispASR backend

app.py

@@ -28,18 +28,35 @@ class LogCapture(io.StringIO):
 # Set up logging
 logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
 
-# Clone and install faster-whisper from GitHub
+# Clone and install dependencies from GitHub
 try:
-    subprocess.run(["git", "clone", "https://github.com/SYSTRAN/faster-whisper.git"], check=True)
-    subprocess.run(["pip", "install", "-e", "./faster-whisper"], check=True)
+    if not os.path.exists("./faster-whisper"):
+        subprocess.run(["git", "clone", "https://github.com/SYSTRAN/faster-whisper.git"], check=True)
+        subprocess.run(["pip", "install", "-e", "./faster-whisper"], check=True)
+    
+    if not os.path.exists("./CrispASR"):
+        logging.info("Cloning CrispASR...")
+        subprocess.run(["git", "clone", "https://github.com/CrispStrobe/CrispASR.git"], check=True)
+        logging.info("Building CrispASR...")
+        os.makedirs("./CrispASR/build", exist_ok=True)
+        subprocess.run([
+            "cmake", "..", 
+            "-DCMAKE_PLATFORM_NO_VERSIONED_SONAME=ON", 
+            "-DWHISPER_BUILD_TESTS=OFF", 
+            "-DWHISPER_BUILD_EXAMPLES=OFF"
+        ], cwd="./CrispASR/build", check=True)
+        subprocess.run(["make", "-j4"], cwd="./CrispASR/build", check=True)
+
 except subprocess.CalledProcessError as e:
-    logging.error(f"Error during faster-whisper installation: {e}")
+    logging.error(f"Error during dependency installation: {e}")
     sys.exit(1)
 
 sys.path.append("./faster-whisper")
+sys.path.append("./CrispASR/python")
 
 from faster_whisper import WhisperModel
 from faster_whisper.transcribe import BatchedInferencePipeline
+import crispasr
 
 # Check for CUDA availability
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
@@ -522,6 +539,8 @@ def get_model_options(pipeline_type):
         return ["cstr/whisper-large-v3-turbo-german-int8_float32","SYSTRAN/faster-whisper-large-v1", "GalaktischeGurke/primeline-whisper-large-v3-german-ct2"]
     elif pipeline_type == "transformers":
         return ["cstr/whisper-large-v3-turbo-german-int8_float32","openai/whisper-large-v3", "openai/whisper-large-v2", "openai/whisper-medium", "openai/whisper-small"]
+    elif pipeline_type == "crispasr":
+        return ["auto", "ggml-base.en.bin", "canary-1b-v2.gguf", "parakeet-tdt-0.6b-v3.gguf", "qwen3-asr-0.6b.gguf"]
     else:
         return []
 
@@ -638,6 +657,19 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                     return_timestamps=True,
                     device=device,
                 )
+            elif pipeline_type == "crispasr":
+                actual_model_path = model_id
+                if model_id == "auto":
+                    # For demo purposes, we'll auto-resolve a decent model if 'auto' is picked
+                    # In a real CLI it would prompt, here we just pick qwen3-asr-0.6b
+                    actual_model_path = "qwen3-asr-0.6b.gguf"
+                
+                # Check if it's a local file, if not try to ensure it's in cache
+                if not os.path.exists(actual_model_path):
+                    logging.info(f"Ensuring model {actual_model_path} is in cache...")
+                    actual_model_path = crispasr.cache_ensure_file(actual_model_path)
+                
+                model_or_pipeline = crispasr.CrispASR(actual_model_path)
             else:
                 error_msg = "Invalid pipeline type"
                 logging.error(error_msg)
@@ -656,6 +688,19 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
         elif pipeline_type == "faster-sequenced":
             segments, info = model_or_pipeline.transcribe(audio_path)
             segments = list(segments)  # Exhaust the generator
+        elif pipeline_type == "crispasr":
+            # crispasr.transcribe returns a list of Segment objects
+            logging.info(f"Transcribing with CrispASR: {audio_path}")
+            raw_segments = model_or_pipeline.transcribe(audio_path)
+            # Map crispasr segments to the format expected by the downstream loop
+            segments = []
+            for s in raw_segments:
+                segments.append({
+                    "start": s.start,
+                    "end": s.end,
+                    "text": s.text,
+                    "timestamp": (s.start, s.end) # for transformers-style compatibility
+                })
         else:
             result = model_or_pipeline(audio_path)
             segments = result["chunks"]
@@ -732,7 +777,7 @@ with gr.Blocks() as iface:
         
         with gr.Row():
             pipeline_type = gr.Dropdown(
-                choices=["faster-batched", "faster-sequenced", "transformers"],
+                choices=["faster-batched", "faster-sequenced", "transformers", "crispasr"],
                 label="Pipeline Type",
                 value="faster-batched"
             )

build.sh

@@ -1,5 +1,17 @@
+#!/bin/bash
+# Clone and install faster-whisper
 echo "downloading faster-whisper from git..."
 git clone https://github.com/SYSTRAN/faster-whisper.git
 cd faster-whisper
-echo "pip installing..."
-pip install .
+pip install .
+cd ..
+
+# Clone and build CrispASR
+echo "downloading CrispASR from git..."
+git clone https://github.com/CrispStrobe/CrispASR.git
+cd CrispASR
+mkdir build
+cd build
+cmake .. -DCMAKE_PLATFORM_NO_VERSIONED_SONAME=ON -DWHISPER_BUILD_TESTS=OFF -DWHISPER_BUILD_EXAMPLES=OFF
+make -j$(nproc)
+cd ../..