Speech & Text Processing Components

AI-powered speech and text utilities unlock voice interfaces, advanced parsing, and smarter context management in your Forjinn workflows. This guide documents built-in Speech-to-Text, Text Splitter, and related modules.

---

Speech-to-Text (STT) Nodes

These nodes convert audio (uploaded or streamed) into usable, searchable text.

Supported Providers

• OpenAI Whisper: Cloud-based, multi-language, high accuracy, used for uploads and real-time
• AssemblyAI, Google STT, etc: Alternative cloud STT when enabled

Inputs

• Audio file (wav, mp3, ogg, etc.) or stream
• Specify language, prompt/context (optional)

Outputs

• Transcribed text
• Optional: word timing, speaker diarization, confidence

Use Cases

• Voice chatbots/IVR
• Meeting/call transcription
• Multimodal data search (voice+text context)

---

Text Splitters

Text Splitter nodes preprocess long documents or user input for chunking—critical for RAG, search, and LLM context building.

Options

• By Characters: Fixed number of chars per chunk
• By Tokens: (preferred for LLMs, e.g., 500 tokens/chunk)
• By Sentence/Paragraph: Use built-in or custom splitter logic
• Overlap: Optional step for retaining context across splits

Use Cases

• Pre-indexing documents for vector search
• Breaking up chat logs, lengthy user submissions, etc

---

Text Processing Utilities

• Normalizer/Cleaner: Whitespace, encoding, character set fix
• Summarizer: Quick auto-summarize for long or multi-part docs
• Keyword Extraction: Highlight or extract key topics/entities
• Regex/Custom Parsers: Build or use provided plugins for special cases

---

Example: Voice-Driven Support Bot

1. User uploads voice memo
2. Speech-to-Text node transcribes to text
3. Text Splitter node (if long)
4. LLM/agent node answers or routes based on transcript

---

Best Practices

• Always preprocess/upload audio in a supported format
• Use correct langs/prompts to improve STT accuracy
• Tune chunking strategy by task (retrieval = small+overlap, doc summarization = longer)
• Chain normalizer before embedding for best search

---

Troubleshooting

• Poor accuracy: Try another provider or adjust input format/preprocessing
• Missing chunks: Lower chunk size, increase overlap, check log for parse errors
• Performance: For long docs/audio, use batch or background jobs to avoid blocking UI

---

Speech and advanced text nodes power truly multimodal, robust, and flexible automations for both human and machine users.