yusasif/intron-stt_tts-benchmark

Code-switched benchmark audio

The exact 16 recordings used to benchmark speech models for the Sahara CodeSwitch Africa Challenge. Published so the reported numbers can be checked against the audio that produced them.

Every clip is intra-sentential code-switching — one speaker moving between a Nigerian language and English inside a single utterance, which is the case the challenge is judged on.

Language	Clips
Hausa	4
Igbo	4
Nigerian Pidgin	4
Yoruba	4

Source

Audio is from intronhealth/AfriSwitch, published by Intron Health. It is reproduced here only because entrants were asked to submit the clips they benchmarked on; all credit for the recordings belongs to Intron. Use of the audio is governed by the terms of the source dataset.

Contents

• data/*.wav — 16 kHz mono
• metadata.csv — file_name, clip_id, language, transcript, source_dataset, and used_in (which results file each clip appears in)

Models benchmarked

ASR: Intron Sahara, ElevenLabs Scribe, Deepgram nova-2-phonecall, and our own fine-tuned STT. TTS: Intron Sahara, ElevenLabs multilingual v2, and our own fine-tuned TTS.

Metrics follow Intron's AfriHealth MultiBench: WER and CER, normalised and unnormalised, reported per language.

Benchmark results

The runs these clips produced are included as JSON, so any number in the report can be traced to the clip and the transcript behind it.

• results/asr_benchmark.json — four ASR models on all 16 clips
• results/asr_sahara.json — the Sahara/Intron ASR run
• results/tts_benchmark.json — TTS round-trip scoring
• results/tts_sahara.json — the Sahara/Intron TTS run

Each row carries the clip id, language, reference, hypothesis, WER and CER (normalised and raw), critical-term recall and latency.

Word error rate, lower is better:

Model	Hausa	Igbo	Pidgin	Yoruba	Overall
Intron / Sahara	0.40	0.55	0.31	0.86	0.53
ElevenLabs Scribe	0.33	0.46	0.28	0.90	0.49
Fine-tuned STT	0.82	0.97	1.14	0.98	0.98
Deepgram nova-2	0.89	0.66	0.66	0.96	0.79

Sahara is not the lowest overall, and it is still the one the live agents run on: it was the only model that kept both halves of a code-switched sentence. On one clip the reference ended "...ni o da o you need to base your" and Sahara returned "...ni ò dá o need to base your", where another model dropped the English tail entirely. A transcript that discards the language the caller actually used cannot drive the right action.

The agent these were benchmarked for

github.com/Yusasif-A/sahara-code-switch — two Nigerian voice agents, a bank fraud line and a telecom care line, that understand English mixed with Pidgin, Yoruba, Hausa or Igbo and answer in the same register.