Voice AI Cost Audit: a demonstration of the audit calculation workflow

This demo uses a different assumed stack from the /voice-example page: its own usage profile, silence share, budget-voice and Cartesia lines, so 'today' here is $0.1389/min LIST x ASSUMED while /voice-example shows $0.123/min LIST x ASSUMED. Both are labelled examples; they are not meant to agree, and neither is a client's bill.

1. Input: agent config and usage

A sample client config (synthetic). The usage profile is stated, not measured, because there is no client here. In a real audit these come from invoices, CDRs and token logs (D1, D2, D4, D6).

# Sample client agent config (synthetic; ASSUMED inputs, not a real client)
agent:
  name: "Riverside / Northwind / Brightline front desk (synthetic)"
  platform: vapi            # Vapi-style orchestration, $0.05/min hosting (LIST)
  architecture: cascaded    # STT -> LLM -> TTS
  language: en-US
  stt:
    provider: deepgram
    model: nova-3           # streaming, regular list $0.0077/min (promo $0.0048 not used)
    billed_on: session_time # silence is billed
  llm:
    provider: openai
    model: gpt-6.1-sol      # $2.00 in / $0.10 cached / $10.00 out per 1M tokens (LIST)
    system_prompt_tokens: 4000        # ASSUMED
    prompt_order: dynamic-first       # ASSUMED: caller/date fields at the top, so prefix caching misses
    turns_per_call: 12                # ASSUMED (3 per minute)
    output_tokens_per_turn: 40        # ASSUMED
  tts:
    provider: elevenlabs
    model: eleven_v4_multilingual     # $0.08 per 1K chars (LIST)
    billed_chars_per_min: 550         # ASSUMED (500 + 10% barge-in waste)
  telephony:
    carrier: twilio
    direction: inbound, US local      # $0.0085/min (LIST), rounded up to the next minute
  silence:
    share_of_connected_time: 0.15     # ASSUMED
    idle_timeout_s: 30                # ASSUMED current setting
    voicemail_detection: off          # ASSUMED
usage:
  connected_minutes_per_month: 100000 # ASSUMED
  avg_call_minutes: 4.0               # ASSUMED
  peak_concurrent_calls: 20           # ASSUMED

2. Workload: 30 synthetic phone calls

Three scripted call types (appointment booking, order status, billing), with names, dates, order numbers, amounts and meter readings drawn from a seeded RNG. Caller turns were synthesized with Piper (open-source CPU TTS, 4 voices: 3 US, 1 UK), band-passed to 300–3,400 Hz, downsampled to 8 kHz mono and mixed with white noise at a random 15–25 dB SNR. Ground-truth text is kept per utterance.

Sample transcript (call00, booking)

AGENT  Thank you for calling Riverside Clinic. How can I help you today?
CALLER Hi, I'd like to book a dental cleaning for next week if possible.
AGENT  Of course. Can I have your full name please?
CALLER Yes, it's Maria Lopez.
AGENT  Thanks. Which day works best for you?
CALLER Would Tuesday at eleven fifteen be available?
AGENT  Yes, Tuesday at eleven fifteen is open. Shall I confirm it?
CALLER Yes please, and could you send me a text reminder the day before?
AGENT  Done. You will get a reminder. Anything else?
CALLER No, that's everything. Thank you so much.

Command and output

$ python scripts/01_make_workload.py --voices /workspace/demo-cache/voices --out workload --calls 30
calls 30 voices ['en_GB-alan-medium', 'en_US-amy-medium', 'en_US-lessac-medium', 'en_US-ryan-medium']
real	0m22.770s
user	1m58.035s
sys	0m6.001s
$ wc -l workload/*.jsonl
  160 workload/manifest.jsonl
   30 workload/transcripts.jsonl

Limit: TTS speech is cleaner and more regular than real callers (no accents beyond 4 voices, no crosstalk, no codec artefacts beyond band-limiting). WER here likely understates what real calls would show.

3. Measure: open STT models, INT8, on CPU

Machine:

$ lscpu (summary)
CPU(s):                                  8
Model name:                              Intel(R) Xeon(R) Processor
Thread(s) per core:                      1
Core(s) per socket:                      8
Socket(s):                               1
L3 cache:                                320 MiB (1 instance)
$ ISA flags of interest
amx_bf16 amx_int8 amx_tile avx512_vnni 
$ grep MemTotal /proc/meminfo
MemTotal:       16386368 kB
$ uname -r
6.12.94+
$ pip versions
ctranslate2        4.8.3
faster-whisper     1.2.1
jiwer              4.0.0
numpy              2.5.4
onnxruntime        1.31.0
piper-tts          1.8.0
scipy              1.18.1
whisper_normalizer 0.1.15
Python 3.13.5

Method: faster-whisper (CTranslate2) with compute_type=int8, greedy decoding, English, no VAD. WER uses jiwer on text passed through the Whisper English normalizer on both sides, so '25194' and 'two five one nine four' match. Pass A: one stream at a time, 8 threads, latency = time from the end of an utterance's audio to its final text. Pass B: W processes x T threads, all 160 utterances split across them and started together, giving throughput (audio seconds processed per wall second) and latency under load.

$ ./scripts/run_stt.sh     # sets ONEDNN_MAX_CPU_ISA=AVX512_CORE_VNNI MKL_ENABLE_INSTRUCTIONS=AVX512, runs 02_bench_stt.py per model
--- results/log_small.en.txt ---
{"model": "small.en", "compute_type": "int8", "decode": "greedy beam=1, language=en", "pass": "A_latency", "threads": 8, "utterances": 160, "audio_s": 497.9, "wall_s": 161.84, "rtf": 0.3250276648183471, "latency_p50_s": 1.0139524534988595, "latency_p95_s": 1.1383606195495304, "latency_max_s": 1.1998456770015764, "wer": 0.019849418206707735, "substitutions": 18, "deletions": 1, "insertions": 10, "ref_words": 1461}
{"model": "small.en", "pass": "B_throughput", "workers": 2, "threads_per_worker": 4, "wall_s": 107.03, "audio_s": 497.9, "throughput_x_realtime": 4.65211337975471, "throughput_x_realtime_per_vcpu": 0.5815141724693388, "latency_p50_s": 1.3106935109990445, "latency_p95_s": 1.518264601451119, "wer": 0.019849418206707735, "substitutions": 18, "deletions": 1, "insertions": 10, "ref_words": 1461}
{"model": "small.en", "pass": "B_throughput", "workers": 8, "threads_per_worker": 1, "wall_s": 81.88, "audio_s": 497.9, "throughput_x_realtime": 6.080995802228017, "throughput_x_realtime_per_vcpu": 0.7601244752785021, "latency_p50_s": 3.868212017501719, "latency_p95_s": 4.377122599201902, "wer": 0.019849418206707735, "substitutions": 18, "deletions": 1, "insertions": 10, "ref_words": 1461}
real	6m5.446s
user	32m7.243s
sys	5m34.615s
--- results/log_large-v3-turbo.txt ---
# Reconstructed 21 Oct 2026 22:27 IST: the original tee capture of this run was empty (stdout block-buffered, job stopped after the 8x1 OOM).
# Lines below are the summary JSON each pass wrote to disk, in the same format 02_bench_stt.py prints.
{"model": "large-v3-turbo", "compute_type": "int8", "decode": "greedy beam=1, language=en", "pass": "A_latency", "threads": 8, "utterances": 160, "audio_s": 497.9, "wall_s": 630.28, "rtf": 1.2658051377763901, "latency_p50_s": 3.965766066499782, "latency_p95_s": 4.551989795200461, "latency_max_s": 4.991652336997504, "wer": 0.01574264202600958, "substitutions": 19, "deletions": 0, "insertions": 4, "ref_words": 1461}
{"model": "large-v3-turbo", "pass": "B_throughput", "workers": 2, "threads_per_worker": 4, "wall_s": 475.68, "audio_s": 497.9, "throughput_x_realtime": 1.0467793647983024, "throughput_x_realtime_per_vcpu": 0.1308474205997878, "latency_p50_s": 5.863953725000101, "latency_p95_s": 6.6741488702002245, "wer": 0.01574264202600958, "substitutions": 19, "deletions": 0, "insertions": 4, "ref_words": 1461}
# 8x1: not completed, OOM (see runs_not_completed.json)
# 4x2: separate run, python 02_bench_stt.py --model large-v3-turbo --workload ../workload --out ../results --workers 4 --skip-a
{"model": "large-v3-turbo", "pass": "B_throughput", "workers": 4, "threads_per_worker": 2, "wall_s": 338.72, "audio_s": 497.9, "throughput_x_realtime": 1.470050918476962, "throughput_x_realtime_per_vcpu": 0.18375636480962024, "latency_p50_s": 8.320900136999626, "latency_p95_s": 9.101172161300928, "wer": 0.01574264202600958, "substitutions": 19, "deletions": 0, "insertions": 4, "ref_words": 1461}
real	5m56.056s
user	38m11.593s
sys	6m56.927s

* Calls per instance = throughput per vCPU x 16 vCPU x 60% max load / 35% caller-speech share (VAD-gated STT only transcribes caller speech). The vCPU equivalence, load cap and speech share are ASSUMED. Calls per instance (16 vCPU) is derived, not measured on a c7i.

Why RTF is high for short turns: Whisper encodes a fixed 30-second window, so a 3-second utterance costs about as much compute as a 30-second one. Per-utterance cost is roughly constant, and the real-time factor on short phone turns is far worse than on long recordings. A streaming model (e.g. Parakeet TDT, or a chunked encoder) would remove this; we did not measure one in this run.

{"model": "small.en", "compute_type": "int8", "decode": "greedy beam=1, language=en", "pass": "A_latency", "threads": 8, "utterances": 160, "audio_s": 497.9, "wall_s": 136.51, "rtf": 0.2741462194726181, "latency_p50_s": 0.8311043414996675, "latency_p95_s": 1.0509468968486542, "latency_max_s": 1.5404997050027305, "wer": 0.044490075290896644, "substitutions": 35, "deletions": 24, "insertions": 6, "ref_words": 1461}
{"model": "small.en", "pass": "B_throughput", "workers": 2, "threads_per_worker": 4, "wall_s": 100.2, "audio_s": 497.9, "throughput_x_realtime": 4.969511993122246, "throughput_x_realtime_per_vcpu": 0.6211889991402807, "latency_p50_s": 1.0853319044981617, "latency_p95_s": 2.2375230923000338, "wer": 0.17796030116358658, "substitutions": 46, "deletions": 199, "insertions": 15, "ref_words": 1461}
{"model": "small.en", "pass": "B_throughput", "workers": 4, "threads_per_worker": 2, "wall_s": 83.32, "audio_s": 497.9, "throughput_x_realtime": 5.976213407822283, "throughput_x_realtime_per_vcpu": 0.7470266759777854, "latency_p50_s": 1.700960041500366, "latency_p95_s": 2.8505559712993964, "wer": 0.21286789869952089, "substitutions": 53, "deletions": 213, "insertions": 45, "ref_words": 1461}
{"model": "small.en", "pass": "B_throughput", "workers": 8, "threads_per_worker": 1, "wall_s": 73.65, "audio_s": 497.9, "throughput_x_realtime": 6.7609373700589055, "throughput_x_realtime_per_vcpu": 0.8451171712573632, "latency_p50_s": 2.969289602000572, "latency_p95_s": 5.360957835998306, "wer": 0.23956194387405885, "substitutions": 57, "deletions": 246, "insertions": 47, "ref_words": 1461}
real	6m53.725s
user	33m40.850s
sys	8m17.416s

4. Quality gates (runbook §7.4, R5 margins)

5. Before: all-in cost per connected minute = $0.1389

Runbook §3.2 split. Silence is moved out of the per-minute lines (platform, STT, telephony) into its own line, so the total is unchanged.

Prices used (all LIST except Cartesia, which is LIST (DERIVED) from the published Scale plan with overage): vapi_platform 0.05 (Vapi hosting, nominal $0.05/min, https://vapi.ai/pricing. After the silence split, the allocated platform line is $0.05 x 85% = $0.0425 (30.6% of this bill); its silent-time share sits in the silence line); deepgram_nova3_stream 0.0077 (regular rate, not the $0.0048 promo, https://deepgram.com/pricing); gpt61sol_in 2.0 ($/1M, https://platform.openai.com/docs/pricing); gpt61sol_cached 0.1 ($/1M); gpt61sol_out 10.0 ($/1M); eleven_v4_per_1k 0.08 (ElevenLabs v4 Multilingual, https://elevenlabs.io/pricing/api); eleven_flash_per_1k 0.04 (ElevenLabs Flash/Turbo, same page); twilio_in 0.0085 (Twilio US local inbound, https://www.twilio.com/en-us/voice/pricing/us); scribe_v2_rt 0.0065 (ElevenLabs Scribe v2 Realtime $0.39/h, https://elevenlabs.io/pricing/api); cartesia_sonic_per_1k 0.03791 (Cartesia Sonic-3.6: NOT a published pay-as-you-go rate. DERIVED from the published Scale plan with overage (1 credit = 1 char): this workload is 100,000 min x 550 chars = 55M chars/month; Scale $299/mo includes 8M credits, the other 47M credits are overage at $38 per 1M = $1,786; $299 + $1,786 = $2,085/month / 55,000K chars = $0.03791/1K. Get a vendor quote before using this in a client refund test. https://www.cartesia.ai/pricing); grok_voice3 0.08 (xAI Grok Voice 3.0 S2S $/min + $0.004/text input, https://docs.x.ai/developers/pricing); voxtral3_agent_large 0.08 (Voxtral 3 agent API with larger model, https://mistral.ai/news/voxtral-3).

6. The three paths

Path 1: tune what runs now (runbook §5)

Caching math: an uncached LLM line resends 13,210 input tokens a minute at $2.00/1M. With a static-first prompt, 51,942 of 52,842 input tokens per call are served from cache at $0.10/1M. Token counts ASSUMED, prices LIST.

Path 2: cheaper API, same tier only (R9)

Speech-to-speech options checked against the STT + LLM + TTS lines they would replace:

Lower tiers were not counted (R9), for example AssemblyAI Universal-Streaming at $0.0025/min or the Voxtral 3 agent API with Mistral Small 4 at $0.04/min, which swaps the LLM tier. They could still be tested from the client's own account.

Path 3: self-host on CPU (runbook §7)

Priced with the deployable config (the lowest p95 under load): small.en, 2 proc x 4 threads. Measured throughput is 0.58x real time per vCPU MEASURED, giving 16.0 calls per c7i.4xlarge. Instances = ceil(20 peak / 16.0) + 1 (N+1) = 3 x $0.714/h LIST x 730 h = $1,564/month MEASURED x LIST x ASSUMED. Invoiced monitoring/support: $0/month ASSUMED (none priced in this demo; idle capacity is already paid because instances are always on). Invoice-comparable: $0.0156/min against Deepgram's $0.0077 LIST; this is the figure used in the 20% test.

Reading: once the CPU is busy, the compute is a fraction of a cent per minute MEASURED x LIST. At 100K min/month, though, average load is about 2.3 calls ASSUMED, and sizing for 20 peak calls needs 2 to 7 always-on instances (N+1) depending on model and config; every config costs more than the API even before internal labour. STT (including its share of silence) is only 5.5% of this bill, so even a self-hosted STT line at $0 would save just 5.5%. large-v3-turbo is about 4x slower than small.en on this CPU, so it needs several times as many instances. On the self-hosted path no audio goes to a model vendor: STT runs on instances the client controls. LLM and TTS self-hosting (e.g. Qwen3-8B, Kokoro-82M on GPU) were not measured and are not priced here. They would also change the LLM tier, so they would need the LLM task check and the listening check before they could count.

Command and output (pricing)

$ python scripts/03_price_paths.py --results results
B = 0.1389 $/min
platform 0.0425
stt 0.00655
llm 0.02762
tts 0.044
telephony 0.00723
silence 0.00993
rounding 0.00106
Path 1: tune current stack S_low 0.0247 (17.8%)  S_high 0.0500 (36.0%)  after 0.1142..0.0889
Path 2: cheaper API, same tier S_low 0.0243 (17.5%)  S_high 0.0243 (17.5%)  after 0.1145..0.1145
Path 3: self-host STT on CPU (LLM, TTS stay on API) S_low -0.0079 (-5.7%)  S_high -0.0079 (-5.7%)  after 0.1468..0.1468
S2S {'option': 'Grok Voice 3.0 ($0.08/min + $0.004/text input)', 'usd_min': 0.08, 'replaces_stt_llm_tts': 0.07932, 'saving': -0.00068}
S2S {'option': 'Voxtral 3 agent, larger model ($0.08/min)', 'usd_min': 0.08, 'replaces_stt_llm_tts': 0.07932, 'saving': -0.00068}
small.en {'wer': 0.0198, 'rtf_single_stream_8t': 0.325, 'p50_s': 1.014, 'p95_s': 1.1384}
    {'workers': 2, 'threads': 4, 'x_realtime_per_vcpu': 0.58151, 'p95_latency_s_under_load': 1.51826, 'wer_under_load': 0.01985, 'calls_per_instance': 15.9501, 'instances': 3, 'compute_usd_month': 1563.66, 'selfhost_stt_usd_min_invoice_comparable': 0.01564, 'selfhost_stt_usd_min_full_cost': 0.01864, 'marginal_usd_min_at_full_load': 0.00045, 'breakeven_min_month_vs_nova3': 550000}
    {'workers': 8, 'threads': 1, 'x_realtime_per_vcpu': 0.76012, 'p95_latency_s_under_load': 4.37712, 'wer_under_load': 0.01985, 'calls_per_instance': 20.84913, 'instances': 2, 'compute_usd_month': 1042.44, 'selfhost_stt_usd_min_invoice_comparable': 0.01042, 'selfhost_stt_usd_min_full_cost': 0.01342, 'marginal_usd_min_at_full_load': 0.00034, 'breakeven_min_month_vs_nova3': 280000}
large-v3-turbo {'wer': 0.0157, 'rtf_single_stream_8t': 1.2658, 'p50_s': 3.9658, 'p95_s': 4.552}
    {'workers': 2, 'threads': 4, 'x_realtime_per_vcpu': 0.13085, 'p95_latency_s_under_load': 6.67415, 'wer_under_load': 0.01574, 'calls_per_instance': 3.58896, 'instances': 7, 'compute_usd_month': 3648.54, 'selfhost_stt_usd_min_invoice_comparable': 0.03649, 'selfhost_stt_usd_min_full_cost': 0.03949, 'marginal_usd_min_at_full_load': 0.00199, 'breakeven_min_month_vs_nova3': None}
    {'workers': 4, 'threads': 2, 'x_realtime_per_vcpu': 0.18376, 'p95_latency_s_under_load': 9.10117, 'wer_under_load': 0.01574, 'calls_per_instance': 5.04017, 'instances': 5, 'compute_usd_month': 2606.1, 'selfhost_stt_usd_min_invoice_comparable': 0.02606, 'selfhost_stt_usd_min_full_cost': 0.02906, 'marginal_usd_min_at_full_load': 0.00142, 'breakeven_min_month_vs_nova3': None}
{
 "baseline_B": 0.1388825,
 "best_candidate_pending_gates": "Path 1: tune current stack",
 "r_low": 0.17764261156013175,
 "r_high": 0.3598829946177524,
 "illustrative_verdict": "straddle (r_low < 20% <= r_high): full refund",
 "status": "illustrative only: Path 1 is the best candidate, pending gates (LLM task-success and listening checks open; bill not verified)",
 "if_both_gates_pass_and_inputs_hold": {
  "r_low": 0.33604989829532156,
  "note": "saving floor would rise to this only if the budget voice (listening check) and the prompt reorder (LLM task-success check) both pass on a client's calls and invoices/logs confirm the cost inputs"
 },
 "path2": "r_high < 20%: full refund",
 "path3": "gates unverified and negative saving: S = 0 (R3)"
}

7. Before vs after, and the recommendation

8. 20% refund test (runbook §9)

Straddle rule: r_low < 20% ≤ r_high gives a full refund. Here the 20% calculation is illustrative and cannot establish a real refund decision: two gates are open (the listening check and the LLM task-success check), and a real audit settles them on the client's calls. The profile here is ASSUMED. In a real audit, the S_low inputs must be MEASURED or LIST, because an ASSUMED number never passes the test on its own.

9. What needs a real client's data

This demo shows the workflow runs and is observable. It cannot produce a client's answer. These inputs only a client can supply:

• Invoices. Net invoices from the platform, STT, LLM, TTS and carrier (credits, discounts, platform margins on pass-through) replace the list-price baseline here. Token logs and CDRs replace the ASSUMED usage profile (tokens per minute, cached share, TTS characters per minute, silence share, call length).
• Real call audio quality. WER must be measured on the client's own held-out calls (real accents, crosstalk, codecs, background noise, their names and numbers), paired against their current STT vendor's WER on the same calls. Synthetic TTS callers likely understate WER.
• Their telephony. The carrier, codec, sample rate, billing increments (per-minute round-up vs per-second) and the end-of-speech to first-audio latency of the current stack (D5) set the rounding line and the +100 ms latency gate.
• Their concurrency. Peak concurrent calls and the hourly curve from CDRs set the self-host instance count and utilisation. Here peak = 20 is ASSUMED; at higher volume or a steadier curve the break-even in §6 could be crossed. The load test should run on the client's cloud or on-prem hardware at their rates.
• The listening check. At least 30 utterances rated by the client's staff decide whether the budget voice (and any Path 2 TTS swap) counts in S_low, and so would decide a real refund verdict (§8 is illustrative only). The LLM task-success check on held-out transcripts settles the prompt reorder.

10. Not measured in this run

• Any paid API (Deepgram, ElevenLabs, OpenAI, Cartesia): none was called in this run and no money was spent, so the current vendors' WER and latency on this audio are unknown.
• Self-hosted LLM and TTS throughput and quality (GPU needed for realistic voice latency).
• A streaming STT model (Parakeet TDT, chunked Whisper). Only batch-on-endpoint Whisper was measured.
• End-to-end turn latency (STT + LLM + TTS + network). Only STT processing latency was measured.
• Performance on an actual c7i.4xlarge. Capacity was scaled from this 8-vCPU VM per vCPU (ASSUMED equivalence).
• Root cause of the AMX int8 accuracy problem.

11. A real run on Modal, 22 October 2026

Two voice-agent turns on one synthetic call, run for real on one rented GPU, to show how a cost gets measured. The same call went through the whole pipeline twice: a first run, then a warm one. It's not a client result, and at this tiny volume most of the bill is start-up.

• Hardware: one NVIDIA T4 on Modal, in one container.
• Speech-to-text: Whisper small.en (MIT licence).
• LLM: Qwen2.5-1.5B-Instruct in fp16, not quantized (Apache-2.0).
• Text-to-speech: Kokoro-82M (Apache-2.0).
• The caller: a 33.9-second billing question, spoken by Kokoro from a written script, so no real person's voice is used.

Start-up, before the first turn: 12.2 s for the container to start, then 3.7 s to load text-to-speech, 2.7 s for speech-to-text and 6.0 s for the LLM.

These two rows cover different things, so read them separately, not as a saving. Modal's bill includes start-up and loading, which a real deployment spreads over many calls, and leaves out idle time between calls; the list-price row is per-use charges only. Neither is a cost per minute. The phone line and any platform fee are left out of both. Modal's figures are from its billing report for app ap-4Ubdw0ij5ZrjsqGkPmQ6bF. The list-price row counts tokens with Qwen's tokenizer, so OpenAI's own count would differ slightly.

What this shows, and what it doesn't

• Quality comes first. The 1.5B model's reply was weak: it asked for the account number the caller had just given, then read it back. A cheaper path has to pass the quality gate, and this one wouldn't.
• Cost depends on utilisation. The warm turn used 4.8 GPU-seconds, about $0.0008 at the T4 list rate, but a GPU that sits idle between calls still bills. What self-hosting costs at volume depends on how busy you keep it. This run makes no savings claim.

Listen to the caller and the agent's reply, or read the script that ran it.

12. Reproduce

All scripts, raw JSON/CSV results and logs are in scripts/ and results/ next to this page (see scripts/README.md). The run is CPU only, on open models. The Piper voices and Whisper weights are downloaded from Hugging Face on first run. Timings vary with machine load; WER was identical across runs with AMX off.

# environment (exact versions: scripts/requirements.lock.txt)
python3 -m venv venv && source venv/bin/activate
pip install -r scripts/requirements.lock.txt
mkdir -p voices && (cd voices && for v in en_US-lessac-medium en_US-ryan-medium en_US-amy-medium en_GB-alan-medium; do python -m piper.download_voices $v; done)
cd scripts
# step 1: workload (seeded: identical transcripts each run)
python 01_make_workload.py --voices ../voices --out ../workload --calls 30
# step 2: STT bench, AMX kernels off (see results/amx_on_finding/)
export ONEDNN_MAX_CPU_ISA=AVX512_CORE_VNNI MKL_ENABLE_INSTRUCTIONS=AVX512
python 02_bench_stt.py --model small.en       --workload ../workload --out ../results --workers 2,8
python 02_bench_stt.py --model large-v3-turbo --workload ../workload --out ../results --workers 2,4
#   (large-v3-turbo with 8 workers needs more than 16 GB RAM; on this box it was OOM-killed)
#   ./run_stt.sh runs the same two models in one go
# step 3 + 4: cost per path and the 20% test
python 03_price_paths.py --results ../results | tee ../results/log_price.txt
# page
python 04_build_html.py --demo ..

Generated by scripts/04_build_html.py. Banner, tags and numbers are produced from results/costs.json. Demo date: 21 Oct 2026.