# OpenRouter ASR capability probe — 2026-09-10

Evidence for `182.C.02.12` (ASR vendor decision) under the rule in
`182.C.01.04`: a claimed provider capability links a dated request/response
fixture and a real artifact.

## What was run

A **round trip with known ground truth**, so the result is a measurement rather
than an impression:

1. `openai/gpt-audio-mini` synthesized speech from a known sentence —
   `"The engine ships on Tuesday and costs forty two dollars."` Streaming is
   mandatory for audio output, and `pcm16` is the only streamed format the
   provider accepts; the raw PCM was wrapped to WAV (24 kHz, mono, 16-bit) with
   ffmpeg. Artifact: `probe.wav`, 3.95 s.
2. That audio was transcribed by three candidates, twice each — once asking for
   a plain transcript, once asking for word-level timestamps as strict JSON.

Full responses: `transcripts.json`.

## Result 1 — transcription is fit for purpose

| Model                              | Transcript                                          | Cost      | Latency |
| ---------------------------------- | --------------------------------------------------- | --------- | ------- |
| `mistralai/voxtral-small-24b-2507` | "The engine ships on Tuesday and costs $42."        | $0.000306 | 0.7 s   |
| `google/gemini-3.1-flash-lite`     | "The engine ships on Tuesday and costs $42."        | $0.000070 | 1.3 s   |
| `openai/gpt-audio-mini`            | "The engine ships on Tuesday and costs 42 dollars." | $0.000066 | 0.7 s   |

All three heard the sentence correctly. **46 audio-input models** are available
on the account, so an ensemble (`182.C.18.07`) is a routing choice rather than a
procurement problem.

### A real defect this surfaced

Two of the three rendered the spoken words "forty two dollars" as **"$42."** —
correct speech, different orthography. The content-word normalizer stripped `$`
as punctuation, lost "dollars", and scored **CWER > 0 on a perfect
transcription**. In exact-dialogue mode that is a blocked delivery caused by the
recognizer's spelling convention.

Fixed in `libs/contracts/src/human-video/dialogue.ts`: currency and percent
symbols are expanded into the words a speaker actually said, before punctuation
is stripped. A second bug fell out of that fix — `"$42."` at the end of a
sentence was being read as a decimal, so the expansion did not fire; the
lookahead now treats `.` as a decimal point only when a digit follows it. Both
are pinned by tests that use these exact measured transcripts.

## Result 2 — word timings are NOT fit for the sync gate

All three returned word-level timestamps when asked. They disagree with each
other on the same 3.95-second audio:

| Pair                                    | Mean onset difference | Worst   |
| --------------------------------------- | --------------------- | ------- |
| voxtral vs gemini-3.1-flash-lite        | 267 ms                | 513 ms  |
| voxtral vs gpt-audio-mini               | 295 ms                | 940 ms  |
| gemini-3.1-flash-lite vs gpt-audio-mini | 562 ms                | 1013 ms |

For scale, the thresholds in `182.C.03.04`:

- ATSC IS-191 production target: **+15 / −45 ms**
- ITU-R BT.1359-1 reject boundary: **+45 / −125 ms**

The worst disagreement is **8.1× the reject boundary**, and even the best mean
disagreement is ~6× the production target. These are _generated_ timestamps from
a chat-shaped model, not forced alignment against the waveform — a model asked
for plausible-looking timings produces plausible-looking timings.

**Conclusion:** they cannot carry `182.C.18.02` (timing stretch, truncation and
re-synthesis detection) or the C1 offset measurement. Using them would produce a
sync verdict whose error is larger than the thing it is measuring.

## Decision recorded for `182.C.02.12`

**ASR vendor: OpenRouter audio-input models. Funded and live** (key active, no
spend limit).

- **Primary:** `mistralai/voxtral-small-24b-2507` — a dedicated speech model.
- **Cheapest:** `google/gemini-3.1-flash-lite` — 4.4× cheaper on this clip and
  equally correct; the default for bulk work per the `CLAUDE.md` cost rule.
- **Ensemble:** available on demand for `182.C.18.07`.

This unblocks the transcription half of `182.C.18`. **Word-level forced
alignment remains blocked** and needs a real aligner — WhisperX,
whisper-timestamped, or Montreal Forced Aligner — on the GPU worker
`182.C.17.13` provisions. That is now the only remaining ASR-side blocker, and
it is `blocked-local`, not `blocked-vendor`.

## Reproducing

`asr_probe.py` (session scratchpad, not committed — it takes the key from the
environment). Total spend for the whole probe: **under $0.002**.
