#!/usr/bin/env python3 """ CASL Voice Bot - Speech Pathology Assistant Main application file that can be used for both local deployment and Hugging Face Spaces """ import os import asyncio import gradio as gr import logging import tempfile import queue import threading import time from dotenv import load_dotenv from openai import AsyncOpenAI # Load environment variables load_dotenv() # Set up logging logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # Initialize OpenAI client with API key from environment openai_client = AsyncOpenAI(api_key=os.getenv("OPENAI_API_KEY")) # Speech Pathologist Agent Prompt SPEECH_PATHOLOGIST_PROMPT = """ You are a speech pathologist, a healthcare professional who specializes in evaluating, diagnosing, and treating communication disorders, including speech, language, cognitive-communication, voice, and fluency disorders. Your role is to help patients improve their speech and communication skills through various therapeutic techniques and exercises. Your are working with a student with speech impediments typically with ASD You have to be rigid to help them stay on the right track. YOu have to start with some sort of intro activity and can not rely on teh student at all to complete your thoughts. You pick a place to start and assess teh speech from there. Each domain from the CASL-2 framework can be analyzed using the sample: Lexical/Semantic Skills: This category focuses on vocabulary knowledge, word meanings, and the ability to use words contextually. It measures both receptive and expressive language abilities related to word use. Key Subtests: Antonyms: Identifying words with opposite meanings. Synonyms: Identifying words with similar meanings. Idiomatic Language: Understanding and interpreting idioms and figurative language. Evaluate vocabulary diversity (type-token ratio). Note word-finding difficulties, incorrect word choices, or over-reliance on fillers (e.g., "like," "stuff"). Assess use of specific vs. vague language (e.g., "car" vs. "sedan"). Syntactic Skills: This category evaluates understanding and use of grammar and sentence structure. It focuses on the ability to comprehend and produce grammatically correct sentences. Key Subtests: Sentence Expression: Producing grammatically correct sentences based on prompts. Grammaticality Judgment: Identifying whether a sentence is grammatically correct. Examine sentence structure for grammatical accuracy. Identify errors in verb tense, subject-verb agreement, or sentence complexity. Note the use of clauses, conjunctions, and varied sentence types. Supralinguistic Skills: This subcategory assesses higher-level language skills that go beyond literal meanings, such as understanding implied meanings, sarcasm, and complex verbal reasoning. Key Subtests: Inferences: Understanding information that is not explicitly stated. Meaning from Context: Deriving meaning from surrounding text or dialogue. Nonliteral Language: Interpreting figurative language, such as metaphors or irony Look for use or understanding of figurative language, idioms, or humor. Assess ability to handle ambiguous or implied meanings in context. Identify advanced language use for abstract or hypothetical ideas. Pragmatic Skills(focus less on this as it is not typically necessary for the age range you will be dealing with): This category measures the ability to use language effectively in social contexts. It evaluates understanding of conversational rules, turn-taking, and adapting communication to different social situations. Key Subtests: Pragmatic Language Test: Assessing appropriateness of responses in social scenarios. Begin by introducing yourself as a speech pathologist and start with a simple vocabulary assessment activity. Be encouraging but structured in your approach. """ # Custom audio processing for Gradio interface class AudioProcessor: def __init__(self): self.transcript = [] self.is_active = False self.voice_model = "shimmer" # Default voice self.notes = [] self.current_assessment = { "lexical_semantic": 0, "syntactic": 0, "supralinguistic": 0, "pragmatic": 0 } async def process_speech(self, audio_data, openai_client): """Process speech using OpenAI's API""" if not self.is_active or audio_data is None: return None, "\n".join(self.transcript), self.get_assessment_html() # Prepare audio file for OpenAI temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".wav") temp_file.close() try: # Save audio data to temporary file sample_rate, audio_array = audio_data import scipy.io.wavfile scipy.io.wavfile.write(temp_file.name, sample_rate, audio_array) # Transcribe audio using OpenAI with open(temp_file.name, "rb") as audio_file: transcript_response = await openai_client.audio.transcriptions.create( file=audio_file, model="whisper-1" ) user_text = transcript_response.text if user_text.strip(): self.transcript.append(f"Student: {user_text}") # Analyze speech for CASL-2 categories self.analyze_speech(user_text) # Generate assistant response chat_response = await openai_client.chat.completions.create( model="gpt-4o", messages=[ {"role": "system", "content": SPEECH_PATHOLOGIST_PROMPT}, {"role": "user", "content": user_text} ] ) assistant_text = chat_response.choices[0].message.content self.transcript.append(f"Speech Pathologist: {assistant_text}") # Generate speech from text speech_response = await openai_client.audio.speech.create( model="tts-1", voice=self.voice_model, input=assistant_text ) # Save speech to temporary file response_temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") response_temp_file.close() speech_response.stream_to_file(response_temp_file.name) # Load audio data for Gradio import soundfile as sf audio_data, sample_rate = sf.read(response_temp_file.name) # Clean up os.unlink(response_temp_file.name) return (sample_rate, audio_data), "\n".join(self.transcript), self.get_assessment_html() except Exception as e: logger.error(f"Error processing speech: {e}") self.transcript.append(f"Error: {str(e)}") finally: # Clean up temp file os.unlink(temp_file.name) return None, "\n".join(self.transcript), self.get_assessment_html() def analyze_speech(self, text): """Analyze speech for CASL-2 categories""" # Simple heuristic analysis - in a real app, this would use more sophisticated NLP # Lexical/Semantic: check vocabulary diversity words = text.lower().split() unique_words = set(words) if len(unique_words) / max(1, len(words)) > 0.7: self.current_assessment["lexical_semantic"] += 1 self.notes.append("Lexical/Semantic: Good vocabulary diversity") # Syntactic: check for sentence complexity sentences = [s.strip() for s in text.replace("!", ".").replace("?", ".").split(".") if s.strip()] avg_words = sum(len(s.split()) for s in sentences) / max(1, len(sentences)) if avg_words > 7: self.current_assessment["syntactic"] += 1 self.notes.append("Syntactic: Complex sentence structures used") # Supralinguistic: check for figurative language (very basic check) figurative_markers = ["like", "as", "than", "seems", "appears", "metaphor", "imagine"] if any(marker in text.lower() for marker in figurative_markers): self.current_assessment["supralinguistic"] += 1 self.notes.append("Supralinguistic: Potential figurative language detected") # Pragmatic: basic check for conversational elements pragmatic_markers = ["hello", "hi", "thanks", "thank you", "please", "excuse me", "sorry"] if any(marker in text.lower() for marker in pragmatic_markers): self.current_assessment["pragmatic"] += 1 self.notes.append("Pragmatic: Appropriate social language detected") def get_assessment_html(self): """Get HTML representation of the current assessment""" html = """

CASL-2 Assessment Progress

""" for category, value in self.current_assessment.items(): category_name = category.replace('_', ' ').title() progress_width = min(100, value * 10) html += f"""
{category_name}
{value} points
""" html += """
""" # Add recent notes if self.notes: html += """

Recent Observations

""" html += "
" return html def start_session(self, voice_model, student_id): """Start a new session""" self.is_active = True self.voice_model = voice_model if voice_model else "shimmer" self.transcript = [] self.notes = [] self.current_assessment = { "lexical_semantic": 0, "syntactic": 0, "supralinguistic": 0, "pragmatic": 0 } student_info = f" for {student_id}" if student_id else "" self.transcript.append(f"Session started{student_info}. The AI Speech Pathologist will speak first.") return "Session active. Please wait for the AI to introduce itself." def stop_session(self): """Stop the current session""" self.is_active = False self.transcript.append("Session ended.") return "Session stopped." def add_note(self, note): """Add a custom note""" if note.strip(): self.notes.append(note) return f"Note added: {note}" return "Note was empty, not added." # Create the audio processor instance audio_processor = AudioProcessor() async def start_session(voice_model, student_id): """Start the speech pathology session""" status = audio_processor.start_session(voice_model, student_id) # Generate initial AI introduction try: # Generate assistant response chat_response = await openai_client.chat.completions.create( model="gpt-4o", messages=[ {"role": "system", "content": SPEECH_PATHOLOGIST_PROMPT}, {"role": "user", "content": "Hello"} # Initial trigger ] ) assistant_text = chat_response.choices[0].message.content audio_processor.transcript.append(f"Speech Pathologist: {assistant_text}") # Generate speech from text speech_response = await openai_client.audio.speech.create( model="tts-1", voice=audio_processor.voice_model, input=assistant_text ) # Save speech to temporary file response_temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") response_temp_file.close() speech_response.stream_to_file(response_temp_file.name) # Load audio data for Gradio import soundfile as sf audio_data, sample_rate = sf.read(response_temp_file.name) # Clean up os.unlink(response_temp_file.name) return status, (sample_rate, audio_data), "\n".join(audio_processor.transcript), audio_processor.get_assessment_html() except Exception as e: logger.error(f"Error starting session: {e}") audio_processor.transcript.append(f"Error: {str(e)}") return status, None, "\n".join(audio_processor.transcript), audio_processor.get_assessment_html() def stop_session(): """Stop the speech pathology session""" return audio_processor.stop_session(), None, "\n".join(audio_processor.transcript), audio_processor.get_assessment_html() async def process_mic_input(audio, progress=gr.Progress()): """Process microphone input""" if audio is None or not audio_processor.is_active: return None, "\n".join(audio_processor.transcript), audio_processor.get_assessment_html() progress(0, desc="Processing speech...") audio_output, transcript, assessment = await audio_processor.process_speech(audio, openai_client) progress(1, desc="Done") return audio_output, transcript, assessment def add_note(note): """Add a note to the session""" result = audio_processor.add_note(note) return "", result, audio_processor.get_assessment_html() def save_session(student_id): """Save session to file""" if not student_id: student_id = "anonymous" # Create session data directory if it doesn't exist os.makedirs("session_data", exist_ok=True) # Save transcript timestamp = time.strftime("%Y%m%d-%H%M%S") filename = f"session_data/{student_id}_{timestamp}.txt" with open(filename, "w") as f: f.write("\n".join(audio_processor.transcript)) f.write("\n\n--- ASSESSMENT NOTES ---\n") for note in audio_processor.notes: f.write(f"- {note}\n") f.write("\n--- CASL-2 SCORES ---\n") for category, score in audio_processor.current_assessment.items(): f.write(f"{category.replace('_', ' ').title()}: {score}\n") return f"Session saved to {filename}" # Create Gradio Interface with gr.Blocks(title="CASL-2 Speech Pathology Assistant") as app: gr.Markdown("# CASL-2 Speech Pathology Assistant") gr.Markdown("### AI-powered speech therapy assessment based on the CASL-2 framework") with gr.Row(): with gr.Column(scale=1): student_id = gr.Textbox(label="Student ID (optional)", placeholder="Enter student ID") voice_select = gr.Dropdown( ["alloy", "echo", "fable", "onyx", "nova", "shimmer"], value="shimmer", label="Assistant Voice" ) start_button = gr.Button("Start Session", variant="primary") stop_button = gr.Button("Stop Session", variant="stop") status = gr.Textbox(label="Status", value="Ready to start") with gr.Accordion("SLP Tools", open=True): note_input = gr.Textbox( label="Add Assessment Note", placeholder="Enter observation or assessment note here..." ) note_button = gr.Button("Add Note") note_status = gr.Textbox(label="Note Status") save_button = gr.Button("Save Session") save_status = gr.Textbox(label="Save Status") with gr.Column(scale=2): audio_output = gr.Audio(label="AI Speech", autoplay=True) audio_input = gr.Audio( label="Speak to the AI", type="microphone", source="microphone", streaming=True ) with gr.Row(): with gr.Column(scale=1): assessment_html = gr.HTML(label="Assessment Progress") with gr.Column(scale=1): transcript = gr.Textbox(label="Transcript", lines=10) with gr.Accordion("About This Application", open=False): gr.Markdown(""" ### About CASL-2 Speech Pathology Assistant This application provides an AI speech pathologist that can assess students using the CASL-2 framework. It focuses on: - **Lexical/Semantic Skills**: Vocabulary knowledge and word usage - **Syntactic Skills**: Grammar and sentence structure - **Supralinguistic Skills**: Higher-level language beyond literal meanings - **Pragmatic Skills**: Social use of language (less emphasis for younger students) The AI will provide structured assessments and exercises to help evaluate speech patterns. ### How to Use 1. Optionally enter a Student ID to track sessions 2. Select the AI voice you prefer 3. Click "Start Session" to begin 4. The AI will introduce itself and begin the assessment 5. Speak into your microphone when it's your turn 6. View the transcript to track the conversation 7. SLPs can add assessment notes as needed 8. Save the session when finished 9. Click "Stop Session" when done ### For Speech-Language Pathologists This tool is designed to supplement, not replace, professional SLP services. SLPs can: - Add custom notes during the session - Save session data for later reference - Track progress across multiple sessions - Use the AI as a consistent assessment tool """) # Setup event handlers start_button.click( fn=lambda voice, student: asyncio.run(start_session(voice, student)), inputs=[voice_select, student_id], outputs=[status, audio_output, transcript, assessment_html] ) stop_button.click( fn=stop_session, outputs=[status, audio_output, transcript, assessment_html] ) note_button.click( fn=add_note, inputs=note_input, outputs=[note_input, note_status, assessment_html] ) save_button.click( fn=save_session, inputs=student_id, outputs=save_status ) # Setup audio processing audio_input.stream( fn=lambda audio: asyncio.run(process_mic_input(audio)), inputs=audio_input, outputs=[audio_output, transcript, assessment_html] ) def main(share=True): """Main function to launch the app""" app.launch(share=share) # Entry point for the application if __name__ == "__main__": main()