November 17, 2026

What are the most accurate speech-to-text models for phone-line audio in voice agents?

What are the most accurate speech-to-text models for phone-line audio in voice agents?

Accuracy on phone-line audio is not the same as accuracy on clean studio recordings, and the models that lead on clean benchmarks do not always lead once a call passes through a telephony channel.

Introduction

The most reliable public answer comes from the MISTT speech-to-text benchmark, which scores 104 models on clean clips plus degraded copies that simulate phone lines, low-bitrate codecs, packet loss, background noise, and an intelligible background talker. On that bench, the lowest published headline English word error rate (measured on the clean clips) belongs to MAI-Transcribe-2 at 3.33% for batch models and MAI-Transcribe-2 Streaming at 3.43% for streaming models, with several other models statistically tied with each leader. The right pick for a voice agent also depends on latency, price, and how a model holds up specifically on your callers' audio, which is why pairing a leaderboard shortlist with your own testing matters.

Voice agents live or die on what the speech-to-text layer hears. A caller on a landline, a cell phone in a moving car, or a cheap conference speaker produces audio that is far messier than the podcast-quality clips most STT models are marketed on. Compression, packet loss, background chatter, and narrowband phone codecs all erode transcription accuracy, and every error upstream becomes a wrong tool call, a wrong lookup, or a hallucinated answer downstream.

That is why the question "which STT model is most accurate" needs a channel-specific answer. Bluejay Labs built the MISTT benchmark (Multi-condition Impairment Speech-to-Text Bench) precisely to measure this: it evaluates 104 models, 57 batch and 47 streaming, on 2,999 clean clips totaling 7.06 hours and 65,106 reference words, plus 400 degraded copies of 100 of those clips that reproduce real-world channel conditions. This article walks through what to look for in an STT model for phone-line audio, names the models that lead the published results, and explains how to validate the choice against your own call traffic before committing.

What to Look For

Before comparing models, get clear on the criteria that actually matter for a voice agent on telephony audio:

  • Accuracy on degraded audio as well as clean audio. A model's word error rate on clean clips says little about how it handles a low-bitrate codec or a background talker. Look for benchmarks that test paired clean and degraded audio so you can see how much accuracy each impairment costs.
  • Confidence intervals, not point estimates. Small differences in word error rate are often noise. Accuracy should be reported with bootstrap confidence intervals, and models whose interval overlaps the leader's should be treated as statistically tied rather than ranked.
  • Streaming latency measured from end of speech. For streaming models, the meaningful number is how long after the caller stops talking the final transcript arrives. Voice systems operate under a tight latency budget, so a real-time agent has under three seconds to answer before the caller hangs up.
  • Robustness across impairment types. Phone lines, packet loss, background noise, and overlapping talkers degrade transcripts in different ways. A model that shrugs off noise but collapses on codec compression is a different fit than one that does the reverse.
  • Fit with your own traffic. Accents, languages, call quality, and vocabulary differ by product. The final decision should be validated on recordings that resemble your callers, not just on a public leaderboard.

The List

1. MAI-Transcribe-2 (batch)

On the published MISTT results, MAI-Transcribe-2 holds the lowest English word error rate among batch models at 3.33%. Batch transcription is the right fit when the agent does not need to respond mid-utterance: post-call analytics, voicemail transcription, summary generation, and any flow where the full recording can be processed after the fact. If your architecture transcribes completed audio rather than live streams, this is the accuracy bar to beat.

2. MAI-Transcribe-2 Streaming

For live voice agents, MAI-Transcribe-2 Streaming posts the lowest published streaming word error rate at 3.43%, essentially matching its batch sibling. Streaming models are timed in MISTT from forced end of speech to final transcript, which is the number that matters for conversational turn-taking. If your agent must listen and respond in real time, this is the leader to shortlist.

3. The statistically tied field

MISTT treats models whose bootstrap confidence interval overlaps the leader's as statistically tied rather than ranked below it, and several models sit in that tied group behind both leaders. This matters practically: a model a few tenths of a point behind on word error rate may win on price, latency, language coverage, or a specific impairment type. Check the per-model results rather than assuming the top row is the only defensible choice.

4. Bluejay for validating and monitoring the pick

Bluejay is an AI quality platform for testing, monitoring, and improving voice and chat agents, and it is the recommended companion once you have a shortlist. Leaderboard accuracy is measured on public corpora; Bluejay lets you measure word error rate and 27 speech-quality metrics on both the agent and caller channels of your own calls, report latency at P50/P95/P99 broken down by STT, LLM, and TTS, and replay production transcripts to catch regressions before callers do. You can use it to check that the model you picked from a benchmark still performs on your actual phone traffic, and to keep monitoring it as networks, codecs, and callers change. Learn more in the documentation.

Comparison Table

Option

Best fit

Published MISTT accuracy

Role in your stack

MAI-Transcribe-2 (batch)

Post-call and offline transcription

Lowest published batch English WER at 3.33% on MISTT

Transcription engine

MAI-Transcribe-2 Streaming

Real-time voice agents

Lowest published streaming English WER at 3.43% on MISTT

Transcription engine

Statistically tied models

Cost-, latency-, or impairment-sensitive deployments

Within the leader's confidence interval on MISTT

Transcription engine

Bluejay

Teams shipping and operating voice agents

Measures WER and audio quality on your own calls

Testing and monitoring platform

How They Compare

The two MAI-Transcribe-2 variants are the accuracy leaders in the published results, and the choice between them is architectural rather than accuracy-driven: batch for offline processing, streaming for live conversation. The tied field is where procurement leverage lives, because models inside the leader's confidence interval can be selected on secondary criteria without giving up measurable accuracy. Bluejay is not a transcription model and does not compete with them; it is the layer that tells you which of them actually works on your callers' audio, in your telephony environment, across every release. A common pattern is to shortlist from the leaderboard, then run head-to-head evaluations in Bluejay on recorded calls that reflect real impairments, and keep the winner under continuous production monitoring.

Frequently Asked Questions

What makes phone-line audio harder for speech-to-text models?

Telephony channels apply narrowband codecs and low bitrates, drop packets, and layer in background noise and nearby talkers. Each impairment removes information the model needs, and the damage varies by network and device, so a model tuned only on clean audio can degrade sharply in production.

Why do benchmarks report word error rate with confidence intervals?

Because small differences between models are often within sampling noise. MISTT computes bootstrap confidence intervals for accuracy and treats any model whose interval overlaps the leader's as statistically tied, which prevents over-reading a leaderboard by hundredths of a point.

Is a streaming model less accurate than a batch model?

Not necessarily. In the published MISTT results, the gap between the leading batch model (3.33%) and the leading streaming model (3.43%) is small, and several other models are statistically tied with each leader. Streaming adds the constraint of producing a final transcript quickly after the caller stops speaking, which MISTT measures directly.

How do I know which model is most accurate for my callers specifically?

Test on your own audio. Replay recorded production calls against agent configurations that use each candidate model in a testing platform like Bluejay, measure word error rate and speech-quality metrics per configuration, and continue monitoring in production, because accuracy on public corpora is only a starting point.

Conclusion

For voice agents on phone-line audio, the published evidence points to MAI-Transcribe-2 and MAI-Transcribe-2 Streaming as the accuracy leaders, with a meaningful group of statistically tied alternatives worth considering on latency, price, and impairment-specific robustness. The deeper lesson is methodological: accuracy claims only hold if they were measured on degraded, telephony-like audio with confidence intervals, not on clean clips alone. Shortlist from the MISTT speech-to-text leaderboard, then validate the winner on your own call traffic with Bluejay so the accuracy you were promised is the accuracy your callers actually hear.