Telephony transcription is speech recognition on phone calls, and phone audio is among the hardest routine input a recogniser sees. Most calls still pass through a network designed for 8 kHz sampling and a 300–3,400 Hz band, often with lossy codecs, packet loss and mobile handsets in noisy places. On top of that, callers read out names, postcodes and account numbers, which are exactly the words a narrowband channel blurs.
Key takeaways
- Telephone audio is typically 8 kHz, band-limited to about 300–3,400 Hz, and encoded with G.711 or a lower-bitrate codec. High-frequency consonant cues are gone.
- Models trained or adapted on telephone speech do markedly better on it than models trained only on wideband audio.
- Record calls dual-channel where possible. Per-channel transcription gives you speaker separation for free.
- Spelled letters and digits are where telephone errors concentrate: "S" and "F", "M" and "N", "fifteen" and "fifty".
- Live calls reach a recogniser through a media fork of the RTP stream; recorded calls arrive as files, often transcoded more than once.
The telephone channel#
The public switched telephone network was built around 8 kHz sampling and G.711 encoding at 64 kbit/s, and much of it still interoperates at that baseline. The usable audio band is roughly 300 to 3,400 Hz. Frequencies above that, where fricatives such as /s/, /f/ and /θ/ carry much of their distinguishing energy, are simply absent. Below 300 Hz, some of the fundamental frequency of lower voices is removed too, although listeners and models can infer pitch from harmonics.
Mobile networks and VoIP add further processing. Mobile calls use codecs such as AMR, which reach good quality at lower bitrates by modelling speech, and are then often transcoded to G.711 at an interconnect. Wideband "HD voice" (AMR-WB, G.722) doubles the bandwidth, but only if every hop in the call supports it; a single narrowband leg reduces the whole call to narrowband. Packet loss and jitter in VoIP produce short dropouts that concealment algorithms paper over, which to a recogniser sounds like smeared or missing syllables.
How call audio reaches a recogniser#
Live calls#
For real-time use, such as agent assist or voice agents on the phone, the telephony platform forks the audio of each call leg and streams it to the recogniser, usually over a WebSocket. The audio typically arrives exactly as the network carried it: 8 kHz μ-law or A-law in 20 ms frames. Sending it through unchanged, and telling the recogniser what it is, avoids a lossy decode and re-encode. Twilio's Media Streams, for example, deliver 8 kHz μ-law audio as base64-encoded messages over a WebSocket.
Recorded calls#
For contact centre analytics, calls are recorded and transcribed afterwards in batch. The key choice is at recording time: dual-channel (agent on one channel, customer on the other) or mixed mono. Dual-channel recordings let you transcribe each side independently, which removes crosstalk from each transcript and gives perfect speaker attribution without diarization. Mono recordings require diarization and suffer wherever both parties speak at once.
Where telephone errors concentrate#
Narrowband audio does not degrade all words equally. Common words in context survive well, because the language model can fill in. Isolated letters, digits and unfamiliar names do not have that support.
| Said | Often heard as | Why |
|---|---|---|
| S | F | Fricative energy above 3.4 kHz is removed |
| M | N | Nasals differ mainly in formant transitions, blurred by the codec |
| B, D, P, T, V | each other | The "E-set" letters share a vowel and differ only in a brief onset |
| fifteen | fifty | Stress and the final /n/ are weak on a lossy channel |
| Van der Berg | Vanderberg, Van de Burg | Rare name, no contextual support |
These are the words that matter most on a service call: account numbers, postcodes, order codes, surnames, amounts and dates. A transcript with a low overall WER can still get most identifiers wrong, so measure entity accuracy separately.
Other telephone-specific factors#
- Non-speech audio. IVR prompts, hold music, ringback tones and DTMF key tones all appear in recordings. Without voice activity detection, some models transcribe them, or hallucinate over them.
- Echo. On mono recordings, and on poorly cancelled lines, the agent's voice can bleed into the customer channel, producing duplicate text.
- Speakerphones and cars. Many calls are made hands-free, adding distance and background noise. See noisy audio.
- Consent and recording law. Call recording and transcription carry legal obligations that vary by country and sector, from informing callers to retaining financial calls. See GDPR and voice data and financial services.
How Reson8 approaches this#
Customer support is one of the industries Reson8 serves, and identifiers read over the phone are the case its Patterns feature targets. Given an expected format such as INV[0-9]{4}[A-Z], the model biases towards matching output: "INB seven six two five X" becomes INV7625X. On Reson8's published evaluation, order number WER fell from 8.0% to 1.5% and licence plate WER from 6.6% to 4.7%. Domain adaptation with text, such as a product catalogue or a list of branch names, applies the same idea to vocabulary.
Common questions#
Why is phone audio worse than a voice memo?
A phone app recording locally captures at 16 kHz or more. A call is squeezed through an 8 kHz, 300–3,400 Hz channel and one or more lossy codecs, which removes much of the high-frequency information that separates similar sounds.
Should I record calls in stereo?
Yes, if your platform supports it. One channel per party makes transcription cleaner and speaker labels exact, and the storage cost is modest.
Do I need a telephony-specific model?
You need a model that has seen plenty of narrowband, conversational audio. Some providers offer separate telephony models; others train one model on mixed data. Test on your own calls either way.
Can a recogniser read DTMF tones?
Speech models are not designed to. DTMF is usually signalled out of band in VoIP (RFC 4733) and handled by the telephony platform, not the recogniser.
References#
- ITU-T Recommendation G.711. Pulse code modulation (PCM) of voice frequencies. International Telecommunication Union.
- Godfrey, J. J., Holliman, E. C., & McDaniel, J. (1992). SWITCHBOARD: Telephone speech corpus for research and development. ICASSP.
- Cieri, C., Miller, D., & Walker, K. (2004). The Fisher Corpus: a Resource for the Next Generations of Speech-to-Text. LREC.
- Schulzrinne, H., Casner, S., Frederick, R., & Jacobson, V. (2003). RTP: A Transport Protocol for Real-Time Applications. RFC 3550, IETF.
- Rosenberg, J. et al. (2002). SIP: Session Initiation Protocol. RFC 3261, IETF.
- Schulzrinne, H., & Taylor, T. (2006). RTP Payload for DTMF Digits, Telephony Tones, and Telephony Signals. RFC 4733, IETF.
- Twilio Media Streams documentation. twilio.com.