November 17, 2026
The most accurate speech-to-text models for voice agents on phone-line audio are not necessarily the ones with the best scores on clean studio recordings.
Introduction
Telephone channels compress audio, drop packets, and layer noise and background talkers over the caller's voice, and those impairments reshuffle the accuracy rankings. Open benchmarks that test models on paired clean and degraded audio, such as the MISTT speech-to-text benchmark, show that MAI-Transcribe-2 and its streaming variant, MAI-Transcribe-2 Streaming, are among the models that lose the least accuracy on the benchmark's Severe phone-line condition.
Most voice agent pipelines start with a transcription step. Whether your stack is a cascaded pipeline (speech-to-text, then a language model, then text-to-speech) or a speech-to-speech model, the agent can only act on what it hears. On a phone call, "what it hears" is a narrowband, lossy, noisy signal, and transcription errors at that stage cascade into every downstream decision: the wrong name in a CRM record, a missed appointment confirmation, a failed verification.
So the question of which speech-to-text model is most accurate on phone-line audio is really a question about how to measure accuracy under realistic channel conditions. This article explains what degrades phone audio, why clean-audio accuracy numbers mislead, how modern benchmarks evaluate robustness, and how to pick a model for your own deployment.
Key Takeaways
- Phone-line audio introduces codec compression, packet loss, background noise, and overlapping talkers, and these impairments change which speech-to-text models rank highest.
- Clean-audio leaderboards are a weak predictor of phone-call performance; robustness should be measured on paired clean and degraded versions of the same recordings.
- MISTT reports MAI-Transcribe-2 at 3.33% English WER among batch models and MAI-Transcribe-2 Streaming at 3.43% among streaming models, and it counts several others in each group as tied rather than ranked below.
- Accuracy should always be reported with confidence intervals; models whose intervals overlap the leader's are tied, not ranked behind it.
- Streaming latency matters as much as accuracy for live calls, so the right choice balances word error rate against how quickly a final transcript arrives after the caller stops speaking.
What Phone-Line Audio Does to Speech
A PSTN phone call is a hostile channel for speech recognition. Traditional phone lines carry narrowband audio, roughly 300 Hz to 3,400 Hz, which discards much of the frequency content that modern models are trained on. Calls routed through VoIP add low-bitrate codecs that further compress the signal, and network congestion introduces packet loss that clips or garbles syllables.
On top of channel impairments, real calls happen in real environments. Callers phone in from cars, busy offices, and kitchens. Background noise competes with the speaker, and a second voice, a TV, or a colleague answering a question in the background can confuse a model that handles only a single clean speaker well.
These impairments do not affect all models equally. A model that is marginally less accurate on clean audio can be substantially more robust to codec artifacts, or the reverse. That is why the ranking you see on a clean benchmark is not the ranking you will experience on a phone call.
Why Clean-Audio Accuracy Numbers Mislead
Most published word error rate (WER) figures come from clean, high-fidelity recordings. Those numbers are useful for comparing models under ideal conditions, but they tell you little about the degraded audio your voice agent will actually receive.
The better approach is paired testing: take the same recordings, keep one copy clean, and produce degraded copies that simulate phone lines, low-bitrate codecs, packet loss, background noise, and an intelligible background talker. Measuring the same model on both versions isolates exactly how much accuracy it loses under each impairment, and which impairment hurts it most. This is the design behind the MISTT speech-to-text benchmark, which evaluates 104 models (57 batch, 47 streaming) on 2,999 clean clips plus 400 degraded copies covering precisely these channel conditions.
Paired testing also surfaces a subtlety that single-number rankings hide: a model can be excellent on clean audio and fall apart on packet loss, while a competitor with a slightly worse clean score degrades far more gracefully. For a voice agent on phone audio, the second model is often the better pick.
How the Top Models Compare
Bluejay Labs publishes MISTT as open research, and its leaderboard reports accuracy with bootstrap confidence intervals rather than bare rankings. That matters because word error rate differences inside the confidence interval are not meaningful; models whose intervals overlap the leader's are treated as statistically tied.
On that leaderboard, the lowest published English WER is 3.33% for the batch model MAI-Transcribe-2, and 3.43% for MAI-Transcribe-2 Streaming, the streaming variant. Several other models are statistically tied with each leader, which means the practical differences among the top tier are smaller than the raw numbers suggest. You can see the full per-model breakdown, including latency and robustness columns, on the MISTT speech-to-text leaderboard.
Two practical lessons follow from this. First, do not pick a model on WER alone when the top several are within statistical noise; look at how each performs specifically on phone-line and codec-degraded audio, and at streaming latency. Second, re-check the leaderboard periodically: the field moves quickly, and today's tie at the top can break either way with the next model release.
Accuracy Is Not the Whole Decision
Even the most accurate model is unusable if it is too slow. A real-time voice agent has a latency budget of under three seconds to respond before the caller hangs up, and transcription is one of the components consuming that budget. Streaming models are timed from the forced end of speech to the final transcript, and that figure belongs next to accuracy in any evaluation.
A sensible selection process looks like this:
- Shortlist on robustness. Filter models by their accuracy on phone-line and codec-degraded audio, not just clean audio.
- Check statistical ties. Treat models inside the leader's confidence interval as equals and differentiate on other axes.
- Measure streaming latency. For live calls, the time from end of speech to final transcript determines how responsive your agent feels.
- Test on your own audio. Benchmarks generalize; your callers' accents, vocabulary, and environments are specific. Run the shortlist on recordings sampled from your production traffic.
That last step is where most teams stall, because building a repeatable evaluation harness is real engineering work. Bluejay is built for exactly this: it tests and monitors voice agents end to end, reports 27 speech-quality metrics, including word error rate, on both the agent and caller channels, and breaks out latency at P50, P95, and P99 by STT, LLM, and TTS so you can see which component is costing you time. If you are evaluating transcription models for a live deployment, Bluejay's documentation walks through how to run those evaluations as automated tests.
Frequently Asked Questions
What makes phone-line audio harder for speech-to-text models than studio audio?
Phone calls combine narrowband frequency response, low-bitrate codec compression, packet loss, background noise, and sometimes an intelligible background talker. Each impairment removes information the model needs, and models degrade by different amounts depending on how they were trained.
Which speech-to-text model has the lowest word error rate on the MISTT benchmark?
The lowest published English WER is 3.33% for MAI-Transcribe-2 in batch mode and 3.43% for MAI-Transcribe-2 Streaming, with several other models statistically tied with each leader based on overlapping confidence intervals.
Should I choose a streaming or batch speech-to-text model for a voice agent?
For live phone calls, you need a streaming model, because the agent must hear and respond in real time. Batch models achieve slightly lower WER in benchmarks but are suited to offline work such as transcribing recorded calls for quality analysis.
When is it worth revisiting which speech-to-text model you use?
Re-evaluate whenever a new model release appears on the leaderboards you follow, and whenever you expand into new caller demographics, accents, or channel types, since robustness profiles differ across conditions.
Conclusion
Accuracy on phone-line audio is a measurement problem before it is a purchasing decision. Clean-audio leaderboards hide exactly the differences that matter on a real call, so look for benchmarks that test paired clean and degraded audio, report confidence intervals, and time streaming latency from end of speech. On current published results, MAI-Transcribe-2 and MAI-Transcribe-2 Streaming lead the MISTT benchmark, with a statistically tied group close behind, but the right model for your agent is the one that holds up on your callers' audio and fits your latency budget.
Once you have shortlisted models, the evaluation does not stop at transcription. The agent's behavior, tool calls, and handoffs all depend on what the model heard, which is why testing the full pipeline matters. Bluejay lets you simulate calls, score speech quality on both channels, and gate deployments on regression tests, so a transcription change never ships unverified. Explore the platform at getbluejay.ai.