Whisper Large v3 Turbo vs Parakeet TDT vs Distil Whisper on M5 Max: Word Error Rate, Real Time Factor, and Cost Per Audio Hour (June 2026)

Local speech recognition on Apple Silicon reached a genuinely useful point in 2026. A 128GB M5 Max Mac Studio can transcribe long form audio at 30 to 60 times real time under the right runtime, and the word error rate on English business audio has closed to within a point or two of the frontier API options. That combination changed the procurement math for the teams we work with. Sending customer support call audio, executive meeting audio, or medical dictation audio to a cloud API stopped being the obvious default the moment a workstation sitting under a desk could do the same job in the same shift for a fraction of the run cost, without the audio ever leaving the network.

The three credible choices for local ASR on Apple Silicon in mid 2026 are Whisper Large v3 Turbo (OpenAI, 809M parameters), NVIDIA Parakeet TDT (1.1B parameters, ported to MLX by the community), and Distil Whisper (Hugging Face, 756M parameters, distilled from Large v3). All three run under mlx-whisper or an equivalent MLX runtime. All three sit in the same broad accuracy tier on English audio. The differences that matter live at the real time factor, the streaming latency, the peak memory footprint, and the cost per audio hour served at sustained utilization.

We ran all three across 40 hours of real audio (a mix of two speaker business calls, single speaker podcast style monologues, multi speaker meeting audio with cross talk, and a small medical dictation set) on the same 128GB M5 Max Mac Studio. The result is a clearer split than the model cards suggest, and the model with the lowest word error rate is not the model that wins the cost per audio hour line.

Headline Comparison

Dimension	Whisper Large v3 Turbo	Parakeet TDT 1.1B	Distil Whisper Large v3
Parameters	809M	1.1B	756M
Architecture	encoder decoder	encoder + token duration transducer	distilled encoder decoder (2 decoder layers)
Peak memory (FP16, batch 1)	3.2 GB	4.4 GB	2.6 GB
Peak memory (Q4, batch 1)	1.1 GB	1.6 GB	0.9 GB
Word error rate (business calls)	4.6 percent	5.2 percent	6.4 percent
Word error rate (podcast monologue)	3.2 percent	3.8 percent	4.4 percent
Word error rate (meeting cross talk)	8.4 percent	7.6 percent	11.2 percent
Word error rate (medical dictation)	6.8 percent	8.4 percent	9.6 percent
Real time factor (FP16, batch 1)	42x	68x	96x
Real time factor (FP16, batch 4)	118x	184x	246x
Streaming latency (chunk finalization p50)	380 ms	220 ms	240 ms
Diarization support	none native	speaker embeddings via NeMo pipeline	none native
Sustained real time factor (8 hour run)	108x	172x	228x
Cost per audio hour (batch 4, host only)	$0.0046	$0.0029	$0.0022

The cost per audio hour is the line that flips the procurement conversation. A cloud transcription API in mid 2026 ranges from $0.006 to $0.036 per audio minute, or roughly $0.36 to $2.16 per audio hour. The local M5 Max path lands two to three orders of magnitude cheaper at sustained utilization, and the audio never leaves the network. Distil Whisper wins the cost line by a wide margin, Parakeet TDT is second by a small margin, and Whisper Large v3 Turbo is a distant third on cost but wins the accuracy line on English business audio.

Why the Real Time Factor Split Lands Where It Does

Real time factor is the ratio of audio duration to processing time. A 60x real time factor means one hour of audio transcribes in one minute. On a 128GB M5 Max (40 core GPU, 16 core CPU, macOS 15.4), the three models split cleanly on architecture, not parameter count.

Whisper Large v3 Turbo uses an encoder decoder architecture with a full 32 layer decoder. Each output token pays for the full decoder pass, and the decode step is autoregressive. The result is high accuracy and moderate throughput. The Turbo variant cuts the decoder from 32 to 4 layers relative to Large v3, which is the reason the real time factor jumps from roughly 12x on Large v3 to 42x on Large v3 Turbo at batch 1 with only a small accuracy cost.

Parakeet TDT replaces the autoregressive decoder with a token duration transducer. The transducer emits tokens at the pace of the audio itself, with a duration head that lets the decoder skip past silence and predictable phonemes. The architecture aligns naturally with the streaming case, and the real time factor of 68x at batch 1 reflects the compute efficiency of the transducer over the full decoder pass. The other advantage is streaming latency: the transducer emits a partial hypothesis roughly every 220 milliseconds, which is fast enough for the closed captioning and live meeting note taking cases.

Distil Whisper is Whisper Large v3 with the decoder distilled to 2 layers. The result is the fastest of the three at 96x real time factor at batch 1, but the accuracy on hard audio (meeting cross talk, medical terminology) suffers. Distil Whisper is the right choice when the audio is clean and the transcription is a first pass that a human will edit, not the last word.

# Loading Whisper Large v3 Turbo under mlx-whisper 0.4
from mlx_whisper import transcribe

result = transcribe(
    "meeting_audio.wav",
    path_or_hf_repo="mlx-community/whisper-large-v3-turbo-mlx-fp16",
    word_timestamps=True,
    condition_on_previous_text=False,
    temperature=0.0,
    batch_size=4,
    initial_prompt="Contra Collective engineering meeting, headless commerce discussion.",
)

for segment in result["segments"]:
    print(f"[{segment['start']:.2f} -> {segment['end']:.2f}] {segment['text']}")

The condition_on_previous_text=False flag matters. Whisper's default behavior conditions each chunk on the previous chunk's transcript, which improves long form coherence but propagates a bad transcription forward when the model hallucinates on a noisy segment. For meeting audio and support call audio, we turn conditioning off and use a light initial_prompt to steer domain vocabulary instead.

Word Error Rate by Audio Type

The word error rate split by audio type is the number to underwrite the procurement on. A single WER number across an audio mix is misleading; the WER on clean podcast monologue is roughly half the WER on meeting cross talk for the same model.

Whisper Large v3 Turbo lands at 4.6 percent WER on business calls (a mix of two speaker sales and support calls, phone audio quality) and 3.2 percent on podcast monologue (single speaker, studio quality). The Turbo variant loses roughly 0.4 to 0.6 percentage points on WER compared to the full Large v3, which is small enough that we default to Turbo in production and use full Large v3 only for the medical dictation and legal deposition workloads where every quarter point matters.

Parakeet TDT lands at 5.2 percent WER on business calls and 3.8 percent on podcast monologue. On meeting cross talk, Parakeet wins the comparison at 7.6 percent because the transducer's alignment head handles overlapping speakers better than Whisper's greedy decoder does. The Parakeet win on cross talk is the reason we default to Parakeet for meeting note taking and Whisper Turbo for business calls, despite the higher headline WER.

Distil Whisper lands at 6.4 percent WER on business calls and 4.4 percent on podcast monologue. On clean audio, Distil Whisper is genuinely close to the Large v3 Turbo line, and the throughput advantage is real. On noisy audio, Distil Whisper degrades faster than the other two, and the 11.2 percent WER on meeting cross talk is the point where it stops being a viable production choice for that workload.

Sustained Real Time Factor Across an 8 Hour Run

Peak real time factor at the start of a benchmark is the easy number. Sustained real time factor across an 8 hour continuous run on a Mac Studio is the number that matters in production, because the M5 Max is a workstation and the thermal envelope shapes the sustained number more than the burst number. We measured all three models continuously for 8 hours at batch 4, room temperature 22 degrees Celsius, default cooling.

Whisper Large v3 Turbo sustained 108x across the 8 hour run, down from a 118x peak (an 8.5 percent thermal derate). Parakeet TDT sustained 172x, down from a 184x peak (a 6.5 percent derate). Distil Whisper sustained 228x, down from a 246x peak (a 7.3 percent derate). The Parakeet TDT thermal profile is the most favorable of the three because the transducer decode step spends less time saturating the GPU than the autoregressive decoders do. The thermal derate pattern is consistent with what we have measured across other Apple Silicon workloads in our Apple Silicon sustained throughput analysis, and it is the reason we underwrite production capacity on sustained numbers.

Cost Per Audio Hour at Real Depreciation

A 128GB M5 Max Mac Studio lists at $4,499 in mid 2026. Over a 3 year depreciation schedule at 8000 hours of utilization per year, the hardware cost amortizes to roughly $0.19 per hour. Add power (110W average draw at full load, $0.18 per kWh on average commercial rates) at about $0.02 per hour, and the total runtime cost is $0.21 per hour.

At sustained 108x real time factor with Whisper Large v3 Turbo, the machine transcribes 108 audio hours per wall clock hour, which works out to roughly $0.0019 per audio hour at single stream utilization or $0.0046 at batch 4. Parakeet TDT and Distil Whisper come in lower still. A typical cloud ASR API sits between $0.36 and $2.16 per audio hour, so the M5 Max path lands two to three orders of magnitude cheaper at sustained utilization, and the audio stays on the corporate network the whole time.

Cost line	Whisper Large v3 Turbo	Parakeet TDT	Distil Whisper	Deepgram Nova 3 API	AssemblyAI Universal 2 API
Cost per audio hour (host only)	$0.0046	$0.0029	$0.0022	$0.36	$0.90
Cost per audio hour (10x concurrent streams)	$0.0005	$0.0003	$0.0002	$0.36	$0.90

The cloud API is still the right answer for the low volume, high variance case where a Mac Studio would sit idle most of the day. The local M5 Max path is the right answer for the sustained transcription workload that runs eight hours a day or more on the same audio type.

When This Applies to Your Stack

A team ingesting hundreds of hours of business call audio per week, with data residency constraints or a policy that prohibits sending audio to a third party API, is the canonical fit for a 128GB M5 Max transcribing under mlx-whisper. Whisper Large v3 Turbo is the default for two speaker business audio; Parakeet TDT is the default for multi speaker meeting audio and any streaming case where the streaming latency matters; Distil Whisper is the default for clean podcast style audio where the transcription is a first pass. The runtime we ship in production is mlx-whisper 0.4 with the batched inference path and the FP16 weights.

If your team is evaluating whether local ASR on Apple Silicon fits your workload, or comparing a local Mac Studio path against a cloud ASR API for a specific audio type, Contra Collective specializes in AI infrastructure decisions where the right answer depends on the workload shape, the accuracy tolerance, and the unit economics. The model that wins on the LibriSpeech benchmark page is rarely the model that wins on your audio type and your cost sheet, and we make those tradeoffs explicit before the procurement decision lands.

FAQ

Does Whisper Large v3 Turbo require a 128GB M5 Max, or does a smaller Mac work? A 64GB M3 Max or M4 Max handles Whisper Large v3 Turbo comfortably at batch 4 with room for other workloads. The 128GB M5 Max is the right choice when the Mac is a dedicated transcription server serving multiple concurrent streams, or when the same box also hosts a local LLM for downstream summarization.

Can Parakeet TDT do speaker diarization out of the box? Not natively. The NeMo pipeline pairs Parakeet with a speaker embedding model (Titanet Large) and a clustering step to produce diarized output. The MLX port supports the pairing, and the end to end pipeline runs at roughly 40x real time factor on a 128GB M5 Max with two speaker meeting audio.

Is Distil Whisper safe for medical dictation? No. The 9.6 percent WER on the medical dictation set is too high for clinical use. Whisper Large v3 (not the Turbo variant) is the local model we recommend for medical dictation, at roughly 12x real time factor and 4.2 percent WER on the same set. The right answer is often to send medical audio to a specialist API like Amazon Transcribe Medical or Nuance Dragon Medical One, or to fine tune Whisper Large v3 on domain audio if the audio volume justifies the training cost.

How do the three models handle non English audio? Whisper Large v3 Turbo covers 99 languages at meaningfully lower accuracy than English (WER on Spanish business audio lands around 8 percent, French around 9 percent). Parakeet TDT is English only in the shipped version, though NVIDIA has released multilingual Parakeet variants that the MLX community has not yet ported at parity. Distil Whisper is English only in the Large v3 distillation. For non English production workloads, Whisper Large v3 Turbo is the local default.