Skip to content

Feedback: speech-to-text-pre-recorded-audio-key-terms-prompting

Original URL: https://assemblyai.com/docs/speech-to-text/pre-recorded-audio/key-terms-prompting
Category: speech-to-text
Generated: 05/08/2025, 4:25:20 pm


Generated: 05/08/2025, 4:25:19 pm

Technical Documentation Analysis: Key Terms Prompting

Section titled “Technical Documentation Analysis: Key Terms Prompting”

This documentation covers a useful feature but has several clarity and completeness issues that could frustrate users. Here’s my detailed feedback:

Problem: The documentation contradicts itself on how keywords are counted.

  • States “up to 1000 domain-specific words or phrases”
  • Later says “Each word in a multi-word phrase counts towards the 1000 keyword limit”
  • Example shows only 3 items but implies complex tokenization

Fix: Provide a clear counting table:

Examples of keyword counting:
• "hypertension" = 1 token
• "differential diagnosis" = 2 tokens
• "Wellbutrin XL 150mg" = 3 tokens
• Total in example: 6 tokens used of 1000 limit

Problem: No guidance on what happens when limits are exceeded or invalid terms are provided.

Add:

  • HTTP error codes and responses for limit exceeded
  • Validation rules for acceptable terms
  • How the system handles rejected keywords

Problem: Claims about “contextual understanding” and “semantic meaning” are unsupported.

Improve: Add a concrete before/after example showing actual transcription improvements.

The page has two H1 headers which creates confusion. Structure should be:

# Key Terms Prompting
## Fine-tuning with keyterms_prompt
## How It Works
## Implementation Examples
## Best Practices
## Troubleshooting

Add upfront:

  • Required API key setup
  • Slam-1 model availability/pricing implications
  • Audio format requirements

Add section:

## Best Practices
- Use domain-specific terminology most likely to be misheard
- Include common abbreviations and acronyms from your field
- Add proper nouns, brand names, and technical terms
- Avoid extremely common words that are rarely mistranscribed

Add domain-specific examples:

  • Medical: drug names, procedures, anatomical terms
  • Legal: case names, legal terminology
  • Technical: product names, technical specifications

Add:

  • What to do if transcription quality doesn’t improve
  • How to identify which terms are actually being recognized
  • Performance impact of using maximum keyword limits

The Python example only checks for HTTP 200 but doesn’t handle:

  • Network timeouts
  • Invalid API keys
  • Malformed responses
  • Transcription service errors

Show what the actual API response looks like, including:

  • Full response schema
  • How keywords affect confidence scores
  • Any metadata about keyword usage
  • Token counter tool to help users estimate their usage
  • Template keyword lists for common industries
  • Success metrics explanation (how to measure improvement)

Add:

  • How this feature works with other AssemblyAI features
  • Performance/cost implications
  • When NOT to use key terms prompting
  1. Fix the counting explanation with a clear formula and examples
  2. Add error response examples with HTTP codes
  3. Include a realistic before/after transcription example
  4. Add input validation rules (special characters, length limits, etc.)
  1. Create a “Quick Start” section with a 3-step process
  2. Add FAQ section addressing common misconceptions
  3. Include performance benchmarks or expected improvement ranges
  4. Add troubleshooting flowchart
  1. Add comprehensive error handling to all code examples
  2. Include response parsing examples showing how to access results
  3. Add validation code for keyword format and limits
  4. Provide async/await alternatives for better JavaScript examples

This documentation would benefit from user testing to identify additional pain points and validation of the technical accuracy of the keyword processing claims.