# Get the proper nouns, codes and numbers right

> Get the proper nouns, codes and numbers right: 4 companies fully do it, 15 with a limit, 41 carry the tag. Every listing in Speech-to-text with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/speech-to-text.names-and-numbers/
- Markdown: https://www.anchorterminal.com/niches/speech-to-text.names-and-numbers/index.md (~1,950 tokens)
- Slim: https://www.anchorterminal.com/niches/speech-to-text.names-and-numbers/index.min.md (~480 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/speech-to-text.names-and-numbers/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 4 | 15 | 41 | 10 | 4 |

## What counts as done

A documented way to pass expected terms before the call.

Measured, for our own tests: each named entity or code in a 50-item test set reproduced character for character in at least 70 percent of cases (the best result in acc-01).

How agents ask for it: transcribe this call, the postcode has to be exact; the customer gives their name and order number, catch them; transcribe this podcast and spell the guests' names correctly

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Google Cloud Speech-to-Text (Google LLC)**: Speech adaptation PhraseSets bias recognition, 100 characters per phrase and 100,000 per request, supported on Chirp 3; no account gives an entity accuracy figure.
- **Ming-3.0-asr-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: Context ASR appends a topic and hotwords to the prompt; the card calls them hints, not a guarantee, and gives no entity accuracy figure.
- **OpenAI Speech to Text (OpenAI OpCo, LLC)**: gpt-transcribe takes prompt and a keywords array for domain terms; no account measures entity accuracy on it.
- **Soniox Speech-to-Text (Soniox Inc.)**: context.terms biases spelling of names and terms; acc-01 measured 71.0 percent exact postcodes on stt-async-v4, the best of seven vendors.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Amazon Transcribe (Amazon Web Services, Inc.)**: Custom vocabularies (50 KB files, Phrase and DisplayAs) apply to batch and streaming, but need a separate create step before the call and no account gives an accuracy figure.
- **AssemblyAI Speech-to-Text (Universal) (AssemblyAI, Inc.)**: keyterms_prompt takes up to 1,000 terms on files and 100 per streaming session; acc-01 measured 66.9 percent exact postcodes on universal-3-pro, under the 70 percent target.
- **Azure AI Speech speech-to-text (Microsoft Corporation)**: Phrase lists bias recognition at runtime, limited to 200 phrases and 3,200 characters for multilingual refinement and about 2,000 at most; no accuracy figure in any account.
- **Cartesia Ink (Cartesia AI, Inc.)**: Realtime keyterms take up to 100 terms totalling 1,200 characters; the batch endpoint has no keyterm parameter and no account measures accuracy.
- **Deepgram Speech-to-Text (Nova-3, Flux) (Deepgram, Inc.)**: keyterm repeats up to 100 terms on Nova-3 and Flux, within a 500-token limit; acc-01 measured 63.7 percent exact postcodes on nova-3, under the 70 percent target.
- **ElevenLabs Scribe Speech to Text API (Eleven Labs Inc.)**: keyterms takes up to 1,000 terms of 50 characters on batch and 50 on realtime, plus a 10,000-term library on Scribe v3 Live. acc-01 ranked ElevenLabs below 71 percent on postcodes without a figure.
- **Gladia Speech-to-Text API + MCP (Gladia SAS)**: custom_vocabulary (beta) takes a list with intensity and pronunciations; no size limit documented and acc-02 placed Gladia below Parakeet and Voxtral on postcodes.
- **Groq Speech-to-Text (Groq LLC)**: prompt gives spelling context within Whisper's 224-token limit; no keyterm list and no account measures accuracy.
- **IBM Watson Speech to Text (International Business Machines Corporation)**: Custom words need a trained custom language model (customization_id), not a per-request list; keyword spotting only reports matches.
- **Inworld STT (Theai, Inc. dba Inworld AI)**: prompts takes an array of names and terms as a soft bias; no list limit or accuracy figure.
- **Mistral Voxtral Transcribe (Mistral AI (RCS Paris 952 418 325))**: context_bias takes up to 100 words or phrases on batch; acc-02 placed Voxtral mid-table on postcodes without a figure.
- **Qwen3-ASR-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Names and terms go in the system message or prompt as context; no limit or accuracy figure is published.
- **Qwen3-ASR-1.7B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Names and terms go in the system message or prompt as context; no limit or accuracy figure is published.
- **Rev AI Speech-to-Text API (Rev.com, Inc.)**: custom_vocabularies phrases on async jobs and custom_vocabulary_id on streaming; no list limit or accuracy figure found.
- **Smallest.ai Pulse STT (Smallest, Inc.)**: keywords takes KEYWORD or KEYWORD:INTENSIFIER entries on Pulse, batch and live; no list limit or accuracy figure found.
- **Speechmatics Speech-to-Text (Cantab Research Ltd)**: additional_vocab with sounds_like, recommended under 1,000 items and rejected above 20,000; acc-01 measured 54.0 percent exact postcodes on enhanced, the lowest of seven.

- not this job: Cohere Transcribe 03-2026 (Cohere Inc.): No prompt, keyterm or context parameter in the docs or processor; acc-01 measured 59.7 percent exact postcodes.
- not this job: Giggy Speech to Text (Giggy): No keyterm, prompt or vocabulary parameter in the speech API docs.
- not this job: Ming-3.0-audio-flash (inclusionAI, Ant Group (named in the inclusion-ai.org footer; no registered company named)): No hotword or context parameter documented for the ASR task; the prompt is a fixed instruction.
- not this job: Ming-3.0-audio-tiny (inclusionAI (Ant Group)): No hotword or context parameter documented for the ASR task.
- not this job: MiniMax Speech-to-Text (Nanonoble Pte. Ltd.): No hotword, prompt or vocabulary parameter in the reference.
- not this job: Nemotron 3.5 ASR Streaming 0.6B (NVIDIA Corporation): No hotword, context or vocabulary input; the only prompt is the language tag.
- not this job: Parakeet TDT 0.6B v3 (NVIDIA Corporation): No hotword, prompt or vocabulary input.
- not this job: Underdog Bark 3.0 (Conway Research): No hotwords, prompt or vocabulary input.
- not this job: Voxtral Mini 4B Realtime 2602 (Mistral AI (RCS Paris 952 418 325)): No context, hotword or prompt input for the weights.

3 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- no context or keyterm parameter (Parakeet, Underdog Bark, Giggy, Groq Whisper beyond the 224-token prompt)
- the only fix is a second LLM pass, which adds a call and a price
- accented postcodes topped out at 71 percent even on the best vendor (acc-01)

## Around it

Before: crm.lookup-expected-names. After: crm.match-record, llm.correct-proper-nouns.

## Where this comes from

Weight 10 of 100 in Speech-to-text: 4 of 32 accounts (acc-01, acc-02, acc-15, acc-05) and the buyer check on names; one account is a 248-clip test across ten vendors

As data: https://www.anchorterminal.com/api/v1/use-cases/speech-to-text.names-and-numbers.json, or the MCP tool get_use_case with id speech-to-text.names-and-numbers. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
