Telnyx STT vs Deepgram vs ElevenLabs on Phone Number Accuracy

Voice agents live or die on transcription accuracy, especially when callers dictate structured data like phone numbers, addresses, or account IDs. A single dropped digit or a misheard word forces the agent to ask the caller to repeat themselves, and that is where conversations fall apart. We built an interactive demo to show how three leading speech-to-text providers handle the same phone number dictation audio clip, side by side.

What is speech-to-text accuracy for voice agents?

Speech-to-text accuracy for voice agents is the rate at which a transcription engine correctly converts spoken audio into text that an agent can act on without clarification. For voice agents, accuracy is not just about word error rate. It also includes entity recognition (phone numbers, dates, names), formatting (dashes, parentheses, country codes), and turn detection (knowing when a caller has finished speaking). A provider with 95% word accuracy but poor entity formatting can still produce unusable output for a voice agent that needs to dial a number or look up an account.

The test: phone number dictation

We recorded a single audio clip of a caller dictating a US phone number with a country code: "My number is plus one, four zero four, five five five, oh one three two." The clip is 8 seconds long, spoken at a natural conversational pace with a slight pause between the country code and the area code.

Phone numbers are a hard STT problem for three reasons:

1. Structured entities: Numbers must be grouped correctly (country code, area code, prefix, line number) and formatted consistently.
2. Digit grouping: Callers often say digits in groups ("four zero four" instead of "four hundred four"), and the engine must reconstruct the intended number.
3. Turn boundaries: The caller may pause mid-number to think, and the engine must hold the turn open until the full number is captured.

Listen to the comparison

The widget below plays the same audio clip through three STT providers in sync. Each row shows the provider's transcription appearing in real time as the audio plays. Click play to hear the audio and watch the transcripts fill in.

Results

Here is what each provider produced from the same audio clip:

Telnyx STT: Captured the full number as "+1 (404) 555-0132" with correct formatting. Held the turn open through the natural pause between the country code and area code, then closed the turn cleanly after the final digit.

Deepgram Flux: Transcribed the number as "+1 four zero four five five five oh one three" (spelled out as words) and dropped the final "two." The turn closed prematurely after "three," cutting off the last digit.

ElevenLabs Scribe v2: Split the utterance into two separate turns. The first turn captured "+1 (404) 555-013" and the second turn captured "2" as a standalone fragment. This forces the agent to either concatenate the fragments or ask the caller to repeat the last digit.

Why turn detection matters

Accuracy is only one third of the problem. Turn detection and latency are the other two. A provider that cuts a turn short forces the agent to interrupt the caller with a clarification prompt, which breaks the conversational flow. A provider that holds the turn open too long creates dead air that makes the caller think the system is broken.

For voice agents, the ideal STT provider delivers three things at once: high entity accuracy, correct turn boundaries, and low end-of-turn latency. Telnyx STT is tuned for all three, with a turn detection model trained on conversational phone audio rather than clean studio recordings.

Try Telnyx STT

Telnyx STT is available now through the Telnyx STT API. You can test it with your own audio clips, integrate it into your voice agent pipeline, and compare it side by side with your current provider. Sign up for a free Telnyx account to get started, or check the STT API documentation for integration details.