> ## Documentation Index
> Fetch the complete documentation index at: https://mumbli-feat-proof-of-use.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Custom Vocabulary for Dictation — Mumbli

> Add proper nouns, brand names, and technical terms to Mumbli's vocabulary. Benchmarked improvement from 36% to 100% accuracy.

# Custom vocabulary for dictation

Speech engines are trained on general language. They struggle with proper nouns, brand names, acronyms, and technical jargon. Mumbli lets you fix that.

## The problem

Without custom vocabulary, speech engines make predictable mistakes:

| You say      | You get        | You wanted |
| ------------ | -------------- | ---------- |
| "VitePress"  | "wheat press"  | VitePress  |
| "ElevenLabs" | "11 labs"      | ElevenLabs |
| "Supabase"   | "super base"   | Supabase   |
| "kubectl"    | "cube control" | kubectl    |

These errors are consistent — the same word gets mistranscribed the same way every time.

## The fix

Open **Settings > Custom Vocabulary** and add the words that matter to you. Mumbli uses them at two levels:

<Steps>
  <Step title="Transcription hints">
    Your vocabulary is sent as a prompt hint to the speech-to-text engine (Groq Whisper). This biases the transcription toward your preferred spellings before any text is generated.
  </Step>

  <Step title="Polishing corrections">
    Your vocabulary is injected into the LLM polishing prompt. Even if the transcription engine still gets it wrong, the polisher knows what you meant and corrects it.
  </Step>
</Steps>

Two layers of correction means errors get caught at both stages of the pipeline.

## Benchmarked results

We tested 11 real vocabulary instances that were consistently mistranscribed:

| Metric                   | Without vocabulary | With vocabulary |
| ------------------------ | ------------------ | --------------- |
| **Exact match accuracy** | 36%                | **100%**        |
| **Test instances**       | 11                 | 11              |

Every single test word was transcribed correctly after adding it to the vocabulary list.

## What to add

Good candidates for your vocabulary list:

* **People's names** — colleagues, contacts, public figures
* **Brand names** — products, companies, services you mention often
* **Technical terms** — frameworks, libraries, CLI tools, protocols
* **Acronyms** — spelled out the way you want them (e.g., "GROQ" not "grock")
* **Domain-specific jargon** — medical terms, legal terms, industry language

## Engine support

| Engine                    | STT hints                 | Polish corrections |
| ------------------------- | ------------------------- | ------------------ |
| **Standard** (ElevenLabs) | Not supported (Scribe v1) | Yes                |
| **Fast** (Groq Whisper)   | Yes                       | Yes                |
| **Deepgram** (Nova-3)     | Not configured            | Yes                |

ElevenLabs Scribe v1 and the current Deepgram path do not use vocabulary hints at the STT level, so corrections happen during polishing only. Groq Whisper supports hints at both levels.

<Card title="Download Mumbli" icon="download" href="https://github.com/fireharp/mumbli/releases/latest">
  Free and open source. Add your vocabulary in Settings after install.
</Card>