A different spelling.
The same word?

Local per-clip transcription with the model already loaded, including audio handling and decoding. All clips included; no warm-up clips excluded. Batch size: 1. Excluded: Model loading, Upload and network time, Gateway overhead.

The GPU was shared with other jobs. These are observed model-call timings, not a service latency promise. API runs include network and provider time, and other local models used different batching, so we do not rank their speed against these numbers.

WER is the sum of substituted, deleted and inserted words divided by the reference word count. CER uses character edits divided by reference characters, excluding whitespace. We aggregate counts across the corpus rather than averaging clip percentages. Insertions can take an error rate above 100%.

Light normalization uses NFKC, lowercase, Arabic vowel-mark and tatweel removal, and unified alef forms. Punctuation is spaced except internal apostrophes and numeric separators. French accents, negation, numbers and word boundaries are retained.

The variant audit permits explicit alternatives, rewrites only accepted hypothesis spans to the reference form, then recomputes WER and CER. The denominators remain 6,990 words and 30,060 characters. These allowances were chosen after inspecting Fennec’s outputs; apply the same frozen policy to all models before making an adjusted ranking.

Exploratory runs on the same audio. OpenAI received generic Tunisian context and Arabic, French and English hints; Gemini received an Arabic hint. No reference transcripts were supplied.

Gemini’s Arabic-hinted result includes one runaway transcript with 2,369 inserted words against a 10-word reference. It remains counted and substantially worsens the corpus score. Two initially failed requests succeeded on retry.

• Blind generalization: these public sets were repeatedly inspected during Fennec development. The overlap check covered the original TunSwitch training manifest, not all training and decoder data.
• Semantic correctness: the variant audit used text, not human adjudication of the recordings. Names, numbers and negation were not separately annotated.
• Isolated speed: local runs shared hardware, while API calls included network time. Qwen used batch 12; its smoke outputs differed from single-clip inference on two of three clips. The final run uses the fixed batch configuration throughout.
• Every possible model setup: Whisper used greedy decoding with Arabic specified. Provider aliases can change, and the commercial APIs did not expose immutable snapshots. No Claude transcription run is included.